Multi-mode optical inspection

Abstract

An inspection system can develop an inspection formula by: using one or more optical inspection subsystems associated with N different optical inspection modes to generate N inspection images of a preliminary sample; using a classifier to generate the probability of each location of the preliminary sample in a background or defect category from at least some combinations of the inspection images from M such optical inspection modes, wherein M is greater than one and less than N and corresponds to the number of such optical inspection modes to be included in the inspection formula; and selecting one of the combinations of the M such optical inspection modes based on a metric describing a difference between the background and defect categories. The inspection system can further use the M inspection images generated using the selected combination of the M such optical inspection modes to identify defects on a test sample.

Description

Technical Field

[0001] The present invention relates generally to defect detection and, more particularly, to defect detection using multiple optical modes. Prior Art

[0002] Inspection systems are commonly used in semiconductor manufacturing processes to identify manufacturing process defects that can lead to performance degradation or failure of a manufactured device. As fabricated feature sizes continue to shrink, the size of manufacturing defects also decreases. This results in weaker measurable signals associated with these defects and lower signal-to-noise ratios (SNRs) during defect detection. Therefore, there is a need to develop systems and methods to address these drawbacks. Summary of the Invention

[0003] According to one or more illustrative embodiments, an inspection system is disclosed. In one illustrative embodiment, the system includes a controller. In another illustrative embodiment, the controller develops an inspection recipe by: using N inspection images of a preliminary sample from one or more optical inspection subsystems associated with N different optical inspection modes; using a classifier to apply the inspection images from at least some combinations of M optical inspection modes to generate a probability that each location of the preliminary sample is in a background or defect class, where M is greater than one and less than N and corresponds to a number of optical inspection modes to be included in the inspection recipe; and selecting one of the combinations of M optical inspection modes based on a metric that describes a distinction between the background and defect classes. In another illustrative embodiment, the controller may use the M inspection images of a test sample generated based on the inspection recipe having the selected combination of M optical inspection modes to identify defects on the test sample.

[0004] According to one or more illustrative embodiments, an inspection system is disclosed. In one illustrative embodiment, the system includes one or more optical inspection subsystems and a controller. In another illustrative embodiment, the controller develops an inspection recipe by: receiving N inspection images of a preliminary sample from the one or more optical inspection subsystems associated with N different optical inspection modes; using a classifier to generate a probability of each of the locations of the preliminary sample being in a background or defect class using the inspection images from at least some combinations of M optical inspection modes, where M is greater than one and less than N and corresponds to a number of optical inspection modes to be included in the inspection recipe; and selecting one of the combinations of M optical inspection modes based on a metric describing a distinction between the background and defect classes. In another illustrative embodiment, the controller may use the M inspection images of a test sample generated based on the inspection recipe having the selected combination of M optical inspection modes to identify defects on the test sample.

[0005] According to one or more illustrative embodiments, an inspection method is disclosed. In one illustrative embodiment, the method includes developing an inspection recipe by: receiving N inspection images of a preliminary sample from one or more optical inspection subsystems associated with N different optical inspection modes; using a classifier to generate a probability that each of the locations of the preliminary sample is in a background or defect class using the inspection images from at least some combinations of M number of the optical inspection modes, where M is greater than one and less than N and corresponds to a number of the optical inspection modes to be included in the inspection recipe; and selecting one of the combinations of M optical inspection modes based on a metric that describes a distinction between the background and defect classes. In another illustrative embodiment, the method includes identifying defects on a test sample using M inspection images of the test sample generated based on the inspection recipe having the selected combination of M optical inspection modes.

[0006] It should be understood that both the foregoing summary and the following detailed description are exemplary and explanatory only and are not necessarily restrictive of the invention as claimed. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and, together with the summary, serve to explain the principles of the invention. Simple diagram description

[0007] Those skilled in the art may better understand several advantages of the present invention by referring to the accompanying drawings.

[0008] FIG. 1A is a block diagram of a detection system according to one or more embodiments of the present invention.

[0009] FIG. 1B is a simplified schematic diagram of an optical imaging subsystem according to one or more embodiments of the present invention.

[0010] FIG. 2A is a flow chart illustrating steps performed in a detection method according to one or more embodiments of the present invention.

[0011] FIG. 2B is a flow chart illustrating steps associated with a detection method according to one or more embodiments of the present invention.

[0012] FIG. 2C is a flow chart illustrating steps associated with a detection method according to one or more embodiments of the present invention.

[0013] 3A illustrates a simulated point cloud associated with background and defect probability for a particular defect type identified by an LDA technique using simulated optical inspection images from two different optical modes generated using a primary mode selected based on SNR, in accordance with one or more embodiments of the present invention.

[0014] FIG3B is a histogram of the probability density functions (PDFs) of background and defect classes along the line in FIG3A according to one or more embodiments of the present invention. For example, the line can be associated with a linear transformation of the point cloud coordinates that best separates the two classes.

[0015] 3C illustrates a simulated point cloud associated with background and defect probabilities for specific defect types identified using simulated optical inspection images from two different optical modes generated by an LDA technique using analysis based on all possible combinations of optical modes, in accordance with one or more embodiments of the present invention.

[0016] FIG. 3D is a histogram of the PDFs of background and defect classes along the line in FIG. 3C , in accordance with one or more embodiments of the present invention.

[0017] 3E is a graph of receiver operating characteristics (ROC) for the multi-mode detection configuration of FIG. 3A-3D and for single-mode operation according to one or more embodiments of the present invention.

[0018] 4 is a graph of captured defects versus disturber points for multi-mode inspection and single-mode inspection using a neural network, according to one or more embodiments of the present invention. Implementation Method

[0019] Cross-reference to related applications

[0020] This application claims the benefit of U.S. Provisional Application No. 63 / 326,268, filed on March 31, 2022, under 35 U.S.C. § 119(e), the entirety of which is incorporated herein by reference.

[0021] Reference will now be made in detail to the disclosed subject matter illustrated in the accompanying drawings. The present invention has been particularly shown and described with respect to certain embodiments and particular features thereof. The embodiments described herein are to be considered illustrative rather than restrictive. Those skilled in the art will readily appreciate that various changes and modifications in form and details may be made without departing from the spirit and scope of the present invention.

[0022] Embodiments of the present invention relate to systems and methods for defect detection using multiple optical modes. As used herein, an optical mode refers to a combination of optical imaging parameters used to generate an image of a sample using an optical inspection tool (e.g., a subsystem), such as, but not limited to, properties of an illumination light beam incident on the sample (e.g., wavelength, polarization, angle of incidence, spot size, spot shape, depth of focus, or the like) or properties of the collected light (e.g., wavelength, polarization, collection angle, or the like). It is contemplated herein that defects on a sample may respond differently to imaging using different optical modes, enabling improved defect analysis or identification by considering images generated using different optical modes. Furthermore, multi-mode inspection can facilitate defect identification with greater sensitivity than a single inspection mode. In this manner, multi-mode inspection can enable identification of defects that are less sensitive and / or less effective than single-mode inspection.

[0023] Semiconductor devices are typically manufactured using a series of process steps, such as, but not limited to, deposition of process layers and / or photoresist, exposure of the photoresist to a desired pattern, etching of the photoresist and / or underlying process layers, polishing, or the like. Variations in any of these processes, alone or in combination, can result in variations in the characteristics of the fabricated samples. Generally speaking, some sample variations may be acceptable and considered upsets, while other sample variations may result in unacceptable performance degradation or device failure and are considered defects of interest (DOIs).

[0024] An inspection tool can be used after a selected process step to inspect a sample for defects. For example, an inspection tool can generate a test image of a sample after a specific process step. This test image is then typically compared to a reference image to identify defects, where the defects are associated with differences between the test and reference images. For example, a reference image can be subtracted from the test image to generate a difference image, where features in the difference image correspond to defects (e.g., sample variation can be measured). These defects can then be classified as DOIs or disturbing points. While various types of inspection tools have been developed, optical inspection tools that generate images using illumination from a light source are commonly used for in-line inspection due to their relatively high processing power.

[0025] Some embodiments of the present invention relate to methods for developing an inspection recipe for optical inspection, wherein the inspection recipe includes a selection of optical modes to be implemented during an inspection process. For example, developing an inspection recipe may include, but is not limited to, generating inspection images of a preliminary sample using multiple optical inspection modes and selecting a subset of these optical inspection modes for use during an inspection process. As an illustration, a series of N inspection images may be generated using N different optical inspection modes, where N is an integer greater than the number of optical inspection modes to be selected for use during the inspection process. A classification technique (e.g., an unsupervised classifier or a supervised classifier) ​​may then be separately applied to various combinations (e.g., subsets) of these inspection images to generate a probability of each sample location represented by a pixel in the inspection image being classified as background or defect based on the respective optical inspection mode. One of these combinations of multiple inspection images (e.g., associated with a combination of optical inspection modes) may then be selected for inclusion in an inspection recipe for inspecting future samples.

[0026] In this way, the process of developing an inspection recipe can identify a selected combination of optical modes from a wide range of possible optical inspection modes that effectively distinguishes defects from background signals. It is contemplated herein that such multi-mode inspection can provide substantially superior performance (e.g., in distinguishing between defects and background signals) compared to single-mode inspection techniques. Furthermore, it can be the case that increasing the number of inspection modes generally increases inspection performance. However, it is further recognized herein that balancing the number of optical inspection modes used during inspection can be desirable, particularly in situations where such multi-mode inspection requires multiple imaging scans and therefore takes longer to execute. Therefore, the systems and methods disclosed herein can be used to identify combinations of optical inspection modes that balance inspection performance and inspection throughput.

[0027] It is contemplated herein that the systems and methods disclosed herein can advantageously identify a combination of optical modes for multi-modal inspection based on complete inspection results generated using the optical modes in question (e.g., a complete set of inspection images comprising raw data from an inspection system). This provides the benefit of multi-modal information at the recipe generation stage and can provide increased performance relative to existing techniques that can provide multi-modal recipe generation based on more limited information.

[0028] For example, some existing techniques may perform a multimodal analysis on a limited dataset containing defects identified by at least one considered optical mode. As an illustration, some existing techniques may perform a complete inspection run in a single mode (e.g., a mode with a relatively high signal-to-noise ratio or other performance metric), identify potential defects based on this single-mode inspection, generate patch images based on these pre-identified potential defects, examine the locations of these pre-identified potential defects using one or more additional modes, and perform a multimodal analysis based on these pre-identified potential defects. However, empirical evidence indicates that DOIs identified using these techniques may have limited success. Specifically, in practice, these techniques may be overwhelmed by nuisances (e.g., background signals) and / or may fail to identify true defects with weak signals because the initial inspection was performed using a single optical mode and is therefore limited by the sensitivity of that single optical mode. Existing multimodal detection technology is generally described in the following cases: U.S. Patent Application Publication No. 2020 / 0025689 published on January 23, 2020, U.S. Patent No. 11,415,531 issued on August 16, 2022, and U.S. Patent No. 11,010,885 issued on May 18, 2021, the entire text of each of which is incorporated herein by reference.

[0029] In contrast, the systems and methods disclosed herein do not require any specific optical mode to identify any specific location as a defect when identifying a combination of optical modes for inclusion in an inspection recipe. Instead, the systems and methods disclosed herein advantageously identify a combination of optical modes based on a multidimensional analysis of complete inspection images from multiple modes. Probability maps can be generated based on a multidimensional analysis of complete inspection images from multiple modes. These probability density maps provide the probability that each location on a sample (e.g., each pixel in aligned images from different modes) may belong to different classes (e.g., background or defect).

[0030] It is contemplated herein that even a complete inspection image using a single optical mode can provide such a large amount of data (particularly for high-resolution inspection) that considerations such as, but not limited to, memory storage or processing speed may impose practical limitations on the ability to simultaneously consider complete inspection images from multiple modes. These considerations may be a factor driving existing multi-mode techniques, such as, but not limited to, multi-mode inspection of defects identified using at least one single-mode technique.

[0031] In some embodiments, complete inspection images associated with all optical modes of interest (e.g., N optical modes) are generated, stored in memory, and then analyzed in combination to determine, based on a performance metric, a subset of the optical modes (e.g., a combination of M optical modes, where M is less than N) to include in an inspection recipe for use during runtime. It should be noted that while M can generally have any value, in some cases it may be desirable to limit M to a relatively small number (e.g., 2 to 5) to balance the performance gains achieved through multimodal inspection with inspection processing power considerations. This technique is applicable, but not limited to, situations where hardware and / or time requirements are acceptable. In some embodiments, various possible combinations of M optical modes are evaluated and compared to each other to determine which combination provides superior performance based on the metric. This technique can reduce practical computational and / or storage issues while maintaining the benefits of a complete multimodal analysis.

[0032] Any suitable metric can be used to select a combination of optical modes for inclusion in an inspection recipe. In some embodiments, the metric is associated with the ability of the combination of optical modes to clearly distinguish between background signals (e.g., nuisance point data) and DOI. As an illustration, the metric can be based on a separation of points associated with different classifications (e.g., background or defect classification) in a multidimensional point cloud.

[0033] Any suitable classifier can be used to provide a probability map within the spirit and scope of the present invention. In some embodiments, the classifier may include an unsupervised and / or supervised classification technique. As one example, a supervised classifier may include a discriminant analysis technique (e.g., a linear discriminant analysis (LDA) technique or the like) applicable to situations where defects can be identified in an optical pattern. As another example, an unsupervised classifier may include a soft clustering technique, such as, but not limited to, Gaussian mixture modeling. For example, a point cloud may be generated using the selected inspection images for each considered combination, where the soft clustering technique generates a probability that each pixel (or associated location on the sample of interest) can be classified into a particular class (e.g., background or defect class). In this way, the metric used to select a combination of optical patterns for inclusion in the inspection recipe can be associated with a separation between clusters associated with background and defects in the point cloud. As another example, an unsupervised classifier includes a neural network. For example, a neural network can be trained to accept inspection images from various optical inspection modalities and generate probabilities for each pixel to be classified as either background or defect. In this way, the output of the neural network can be of substantially the same type as other techniques (e.g., soft clustering based on point clouds, or similar). A neural network can also be trained to accept additional information. For example, a neural network can be trained to accept design data (e.g., associated with patterned features on a sample) so that spatial relationships between pixels can be considered when determining the probability that a particular pixel (and therefore the corresponding location on the sample) is classified as background or defect.

[0034] Additional embodiments of the present invention relate to methods of performing inspection using an inspection recipe once a combination of optical patterns has been identified.

[0035] Additional embodiments of the present invention relate to inspection systems suitable for performing multi-mode optical inspection.

[0036] 1A to 4 , a system and method for multi-mode optical inspection will be described in more detail according to one or more embodiments of the present invention.

[0037] FIG. 1A is a block diagram of a detection system 100 according to one or more embodiments of the present invention.

[0038] In one embodiment, the detection system 100 includes one or more optical imaging subsystems 102 (e.g., optical imaging tools) configured to generate one or more images of a sample 104, wherein the one or more optical imaging subsystems 102 can be configured to use multiple optical detection modes to image the sample 104. For example, an optical imaging subsystem 102 can include an illumination subsystem 106 configured to illuminate the sample 104 using illumination 108 from an illumination source 110 and a light collection subsystem 112 configured to generate an image of the sample 104 using a detector 116 in response to light emitted from the sample (e.g., sample light 114) and illumination 108.

[0039] An optical detection mode may correspond to any combination of parameters associated with imaging sample 104, including, but not limited to, the properties of illumination directed to sample 104 or light collected from sample 104. Optical detection modes are described in more detail below.

[0040] Sample 104 may include a substrate (e.g., a wafer or the like) formed from a semiconductor or non-semiconductor material. For example, a semiconductor or non-semiconductor material may include, but is not limited to, single crystal silicon, gallium arsenide, and indium phosphide. Sample 104 may further include one or more layers disposed on the substrate. For example, these layers may include, but are not limited to, a resist, a dielectric material, a conductive material, and / or a semiconducting material. Many different types of such layers are known in the art, and the term sample, as used herein, is intended to encompass a sample on which all types of such layers may be formed. The one or more layers formed on a sample 104 may be patterned or unpatterned. For example, a sample may include a plurality of dies, each having repeatable patterned features. The formation and processing of these material layers may ultimately result in a completed device. Many different types of devices may be formed on a sample, and the term sample, as used herein, is intended to encompass a sample on which any type of device known in the art is fabricated.

[0041] Optical imaging subsystem 102 can use any technique known in the art to generate one or more images of sample 104. In some embodiments, optical imaging subsystem 102 is an optical imaging subsystem 102 in which illumination source 110 is an optical source configured to generate illumination 108 in the form of light, and in which light collection subsystem 112 images sample 104 based on light emitted from sample 104.

[0042] Furthermore, imaging using different optical detection modes can generally be performed using any number of optical imaging subsystems 102. In some embodiments, a single optical imaging subsystem 102 can be configured to image the sample 104 using multiple optical detection modes simultaneously or sequentially. In some embodiments, different optical imaging subsystems 102 are used to provide at least some of the different optical detection modes.

[0043] FIG. 1B is a simplified schematic diagram of an optical imaging subsystem 102 according to one or more embodiments of the present invention.

[0044] Illumination source 110 may include any type of illumination source known in the art suitable for generating optical illumination 108, which may be in the form of one or more illumination light beams. Furthermore, illumination 108 may have any spectrum, such as, but not limited to, extreme ultraviolet (EUV) wavelengths, ultraviolet (UV) wavelengths, visible wavelengths, or infrared (IR) wavelengths. Furthermore, illumination source 110 may be a broadband source, a narrowband source, and / or a tunable source.

[0045] In some embodiments, illumination source 110 comprises a broadband plasma (BBP) illumination source. In this regard, illumination 108 may comprise radiation emitted by a plasma. For example, a BBP illumination source 110 may (but need not) comprise one or more pump sources (e.g., one or more lasers) configured to focus into a volume of gas, thereby causing energy absorption by the gas to generate or maintain a plasma suitable for emitting radiation. Furthermore, at least a portion of the plasma radiation may be used as illumination 108.

[0046] In another embodiment, the illumination source 110 may include one or more lasers. For example, the illumination source 110 may include any laser system known in the art that is capable of emitting radiation in the infrared, visible, or ultraviolet portions of the electromagnetic spectrum.

[0047] The illumination source 110 can further generate illumination 108 having any temporal profile. For example, the illumination source 110 can generate continuous wave (CW) illumination 108, pulsed illumination 108, or modulated illumination 108. Additionally, the illumination 108 can be delivered from the illumination source 110 via free-space propagation or guided light (e.g., an optical fiber, a light pipe, or the like).

[0048] The illumination subsystem 106 and / or the optical imaging subsystem 102 may include various components for directing illumination 108 toward the sample 104, such as, but not limited to, lenses 118, mirrors, or the like. Furthermore, these components may be reflective or transmissive. In this manner, the depiction of lens 118 as a transmissive element in FIG. 1B is merely illustrative and non-limiting. The illumination subsystem 106 may further include one or more optical elements 120 for modifying and / or conditioning light in the associated optical path, such as, but not limited to, one or more polarizers, one or more filters, one or more beam splitters, one or more diffusers, one or more homogenizers, one or more apodizers, or one or more beam shapers.

[0049] In some embodiments, the inspection system 100 includes a translation stage 122 for holding and / or positioning the sample 104 during imaging. For example, the translation stage 122 may include any combination of linear actuators, rotary actuators, or angular actuators to position the sample 104 using any number of degrees of freedom.

[0050] The optical imaging subsystem 102 may include various components for collecting at least a portion of the sample light 114 radiation (e.g., in the case of an optical imaging subsystem 102, sample light) emitted from the sample 104 and directing at least a portion of the sample light to a detector 116 for generating an image. An image generated by the detection system 100 may be any type of image known in the art, such as, but not limited to, a brightfield image, a darkfield image, a phase contrast image, or the like. Furthermore, the images may be stitched together to form a composite image of the sample 104 or a portion thereof.

[0051] Inspection system 100 can further image sample 104 using any technique known in the art. In some embodiments, inspection system 100 generates an image of sample 104 in a scanning mode by focusing illumination 108 onto sample 104 as a point or line of light, capturing a point or line image, and scanning sample 104 to create a two-dimensional image. In this configuration, scanning can be achieved by moving sample 104 relative to illumination 108 (e.g., using translation stage 122), by moving illumination 108 relative to sample 104 (e.g., using an actuable mirror or the like), or a combination thereof. In some embodiments, inspection system 100 generates an image of sample 104 in a static mode by directing illumination 108 toward sample 104 in a two-dimensional field of view and capturing the two-dimensional image directly using detector 116.

[0052] The optical imaging subsystem 102 may include various components for directing the sample light 114 to the detector 116, such as, but not limited to, lenses 124, mirrors, or the like. Furthermore, these components may be reflective or transmissive. In this manner, the depiction of lens 118 as a transmissive element in FIG. 1B is merely illustrative and non-limiting. The optical imaging subsystem 102 may further include one or more optical elements 126 for modifying and / or conditioning light in the associated optical path, such as, but not limited to, one or more polarizers, one or more filters, one or more beam splitters, one or more diffusers, one or more homogenizers, one or more apodizers, or one or more beam shapers.

[0053] Detector 116 may include any type of sensor known in the art suitable for measuring sample light 114. For example, detector 116 may include a multi-pixel sensor, such as, but not limited to, a charge-coupled device (CCD), a complementary metal-oxide semiconductor (CMOS) device, a line sensor, or a time-delayed integration (TDI) sensor. As another example, a detector 116 may include two or more single-pixel sensors, such as, but not limited to, a photodiode, a collapsed photodiode, a photomultiplier tube, or a single-photon detector.

[0054] The illumination subsystem 106 and the optical imaging subsystem 102 can be configured in various ways within the spirit and scope of the present invention. In some embodiments, as shown in FIG. 1B , the detection system 100 includes at least one beam splitter 128 that is common to the optical paths of the illumination subsystem 106 and the optical imaging subsystem 102. In this manner, the illumination subsystem 106 and the optical imaging subsystem 102 can both share a common objective lens 130 and utilize the entire available pupil or numerical aperture (NA) provided by the objective lens 130. In some embodiments, although not shown, the illumination subsystem 106 and the optical imaging subsystem 102 can have separate optical paths without common elements.

[0055] In some embodiments, various alignment and / or stitching operations are performed on the data generated by optical imaging subsystem 102 to form one or more images of sample 104. As an example, in the case of a scanning system, data associated with multiple swaths or portions thereof can be aligned and / or stitched to form an image of the entire sample 104 or only a portion thereof. In this manner, the term image is used herein to broadly describe a pixel array representing a portion of sample 104 and is not intended to confer limitations associated with imaging techniques. Thus, an image can correspond to a complete data set provided by optical imaging subsystem 102, a subset of this data (e.g., a sub-image), or multiple data sets that have been stitched (and appropriately aligned).

[0056] In some embodiments, the inspection system 100 includes a controller 132 communicatively coupled to the optical imaging subsystem 102 or any component thereof. The controller 132 may include one or more processors 134 configured to execute program instructions maintained on a memory 136 (e.g., a memory medium). In this regard, the one or more processors 134 of the controller 132 may perform any of the various procedural steps described throughout the present invention.

[0057] The one or more processors 134 of a controller 132 can include any processing element known in the art. In this sense, the one or more processors 134 can include any microprocessor-type device configured to execute algorithms and / or instructions. In one embodiment, the one or more processors 134 can be comprised of a desktop computer, a mainframe computer system, a workstation, a video computer, a parallel processor, or any other computer system (e.g., a network-connected computer) configured to execute a program configured to operate the detection system 100, as described throughout this disclosure. It should be further understood that the term "processor" can be broadly defined to encompass any device having one or more processing elements that execute program instructions from a non-transitory memory 136.

[0058] Memory 136 may comprise any storage medium known in the art suitable for storing program instructions executable by the associated processor(s) 134. For example, memory 136 may comprise a non-transitory memory medium. By way of another example, memory 136 may include, but is not limited to, a read-only memory, a random access memory, a magnetic or optical memory device (e.g., a disk), a tape, a solid-state drive, and the like. It should be further noted that memory 136 may be housed in a common controller housing with one or more processors 134. In one embodiment, memory 136 may be remotely located relative to the physical location of one or more processors 134 and controller 132. For example, one or more processors 134 of controller 132 may access a remote memory (e.g., a server) accessible via a network (e.g., the Internet, an intranet, and the like). Therefore, the above description should not be construed as limiting the present invention and is merely illustrative.

[0059] An optical detection mode provided by the optical imaging subsystem 102 may be controlled based on control of any combination of components in the illumination subsystem 106 or the light collection subsystem 112. For example, control of the illumination 108 directed to the sample 104 may be provided directly by the illumination source 110 and / or by optical elements 120, such as, but not limited to, a spectral filter for controlling a wavelength of the illumination 108, a polarizer for controlling a polarization of the illumination 108, or an apodizer for controlling an angular distribution of the illumination 108 on the sample 104 (e.g., in an illumination pupil plane). As another example, control of the sample light 114 collected from the sample 104 and passed to the detector 116 may be provided by the optical element 120, such as, but not limited to, a spectral filter for controlling a wavelength of the sample light 114 passed to the detector 116, a polarizer for controlling a polarization of the sample light 114 passed to the detector 116, or an apodizer for controlling an angular distribution of the sample light 114 passed to the detector 116 (e.g., in a collection pupil plane).

[0060] As an illustration, a particular optical detection mode may correspond to illumination 108 having a selected spectrum (e.g., as described by a bandwidth and / or a central wavelength) and a selected polarization directed onto the sample at a selected angle of incidence (e.g., as defined by an illumination aperture or apodizer). The particular optical detection mode may further correspond to a particular spectrum and polarization of sample light 114 directed to detector 116 (both of which may be the same as or different from illumination 108 incident on sample 104).

[0061] Furthermore, either illumination source 110 and / or optical element 120 can be adjusted so that detection system 100 can be configured to provide different optical detection modes. For example, either optical element 120 can be directly tuned and / or controlled by an actuator to provide different optical detection modes. In some embodiments, controller 132 generates drive signals for either illumination source 110 and / or optical element 120 to selectively provide different optical detection modes.

[0062] In some embodiments, the detection system 100 sequentially provides images using different optical detection modes. For example, the detection system 100 can sequentially switch between different optical imaging subsystems 102 and / or adjust parameters of an optical imaging subsystem 102 to provide sequential images of the sample 104 using different optical detection modes. As another example, an optical imaging subsystem 102 can be configured to simultaneously provide two or more images using different optical detection modes. As an illustration, an optical imaging subsystem 102 can include two or more collection channels, each having a separate detector 116. The optical imaging subsystem 102 can then include additional optical elements 120 (e.g., separate spectral filters, polarizers, or the like) for splitting the sample light 114 into the various channels and / or for providing separate control over the properties of the sample light 114 directed to the associated detector 116 in each channel.

[0063] FIG2A is a flow chart illustrating the steps performed in a testing method 200 according to one or more embodiments of the present invention. Applicants note that the embodiments and implementation techniques previously described herein in the context of testing system 100 should be interpreted as extending to method 200. However, it should be further noted that method 200 is not limited to the architecture of testing system 100. In some embodiments, controller 132 is configured to develop a testing recipe using method 200 or any steps thereof.

[0064] In some embodiments, method 200 includes a step 202 of developing an inspection recipe. In some embodiments, method 200 includes a step 204 of identifying defects on a test sample using the inspection recipe.

[0065] FIG2B is a flowchart illustrating steps (e.g., sub-steps) associated with step 202 of developing a test recipe according to one or more embodiments of the present invention. FIG2C is a flowchart illustrating steps (e.g., sub-steps) associated with step 204 of developing a test recipe according to one or more embodiments of the present invention.

[0066] In some embodiments, step 202 of developing an inspection recipe includes step 206 of generating N inspection images of a preliminary sample using one or more optical imaging subsystems 102 associated with N different optical inspection modes. As previously described herein, an optical inspection mode can include any unique combination of optical imaging parameters of one or more optical imaging subsystems 102. For example, an optical inspection mode can include a unique combination of parameters associated with illumination 108 and / or sample light 114 used to generate an image, such as, but not limited to, wavelength, polarization, or angle.

[0067] Step 206 may include generating any number of detection images. In some embodiments, N is an integer greater than 2. In some embodiments, N is an integer greater than 5, 10, 100, or greater. For example, N may be greater than or equal to 10.

[0068] The N test images can be any type of image, including but not limited to darkfield or brightfield images. In some embodiments, the test image can be a raw image from an optical imaging subsystem 102. In this configuration, the test image can include various patterned features on the sample 104. In some embodiments, the test image is a difference image. For example, a difference image can be generated based on a difference (e.g., a subtraction) between a raw image and a reference image that can represent an ideal or reference. In this configuration, a test image can represent the deviation of a raw image of a test area of ​​interest from the reference. This reference image can be generated using any suitable technique. For example, a reference image can be generated using a single image of a reference area that has a common design as the test area and is known or expected to be defect-free. As another example, a reference image can be generated by combining (e.g., averaging, etc.) multiple images of multiple reference areas.

[0069] In some embodiments, step 206 further includes registering and / or scaling the N detection images. In this way, the N detection images may have a common number of pixels and each pixel of all the detection images may correspond to a common position on the preliminary sample.

[0070] In some embodiments, step 202 of developing an inspection recipe includes step 208 of using a classifier (e.g., an unsupervised and / or a supervised classifier) ​​to utilize inspection images from at least some combinations of a number M of optical inspection modes to generate a probability for each location of the preliminary sample to be in a background or defect class. Furthermore, M can be an integer greater than one and less than N and corresponds to the number of optical inspection modes to be included in the inspection recipe. As used herein, a background class may correspond to a class that describes background signals (e.g., disturbers) that are not of interest during inspection, while a defect class describes outliers or rare events that may be associated with physical variations on the sample. It should be noted that defects in this class may generally include defects of interest (DOIs) that may affect device performance, as well as additional disturber signals that may not generally affect device performance.

[0071] Furthermore, it should be noted that a classifier can generally provide probabilities associated with any number of classes, where the number of classes can be fixed or dynamically generated. For example, it may be the case that different types of defects on a sample 104 can be separately identified (e.g., differentiated) by the classifier. In this manner, the classifier can generate probabilities associated with a single defect class or multiple defect classes that describe all defects. For example, defect classes may correspond to (but are not limited to) different defect types, defect locations, or performance impacts.

[0072] The process of generating the probability of each location in the preliminary sample being in the background or defect class can include generating a series of probability maps or probability density functions (PDFs) for each class. As an example, a PDF for a background class provides a probability for each pixel in the N inspection images (corresponding to each associated location on the preliminary sample) to be in the background class, while a PDF for a defect class provides a probability for each pixel (and each corresponding location) to be in the defect class. Furthermore, it is contemplated herein that an unsupervised technique can generally characterize data into any number of classes (e.g., a background class and any number of defect classes).

[0073] In some embodiments, step 202 of developing an inspection recipe includes step 210 of selecting one of M combinations of optical inspection modes as a selected combination based on a metric describing a distinction between background and defect classes. In this manner, steps 208 and 210 can select a combination of optical inspection modes (e.g., a set of optical inspection modes) suitable for distinguishing between background and defect signals based on the metric. For example, a receiver operating characteristic (ROC) curve can be generated based on a threshold (e.g., a probability threshold) for a detection classification. The area under the curve can be used as a scalar metric to rank mode combinations, where a larger area indicates a more accurate result. As another example, the metric can be associated with the misclassification error of known test data at a fixed true positive rate.

[0074] In other words, step 208 may include executing the classifier (or different iterations of the classifier) ​​multiple times using inspection images associated with different combinations of optical inspection modes. Step 210 may then include selecting one of the analyzed combinations for inclusion in the inspection recipe based on its ability to distinguish defects from background signals.

[0075] In a general sense, the total number of possible combinations of M optical detection modes from a set of N possible optical detection modes can be described by the binomial coefficients given in equation (1): (1)

[0076] As a simple illustration, applying equation (1) with values ​​of M = 2 and N = 50 results in a total of 1,225 possible combinations. Therefore, it is contemplated herein that the number of possible combinations may become prohibitive in practice, and a variety of factors may be considered when selecting the values ​​of M, N, and the number of possible combinations considered in step 208.

[0077] For example, M corresponds to the number of optical inspection modes to be included in the inspection recipe and implemented each time the inspection recipe is run. The value of M can be influenced by factors such as, but not limited to, the inspection processing power tolerances for the specific application and the specific optical imaging subsystem(s) 102 utilized. As previously described herein, some optical imaging subsystems 102 may be capable of generating images using multiple optical inspection modes simultaneously (e.g., via different collection channels), while some optical imaging subsystems 102 may require multiple separate measurements (e.g., scans) to acquire images using different optical inspection modes. Therefore, the number M can (but need not) be selected to balance processing power with the performance increase achieved through multi-mode inspection. In some embodiments, the value of M is two. In some embodiments, the value of M is three. In some embodiments, the value of M is in the range of 3 to 5. In some embodiments, the value of M is greater than 5.

[0078] Furthermore, the value of N corresponds to the number of optical inspection modes considered during development of an inspection recipe (e.g., step 202). In this manner, the value of N may be influenced by factors such as (but not limited to) the physical limitations of an optical imaging subsystem 102, the time requirements associated with inspection recipe development, or the computational requirements associated with inspection recipe development. For example, it may be possible to obtain a number of optical modes associated with different narrowband wavelength spectra using broadband illumination 108 from a broadband illumination source 110 (e.g., a BBP source or the like). Furthermore, additional variations in polarization, angle of incidence, or the like may further increase the possible number of optical inspection modes theoretically available from an optical imaging subsystem 102. However, it may be possible that time or computational requirements may practically limit the number N selected when implementing method 200 in a particular application.

[0079] Furthermore, not all possible combinations of M optical inspection modes need to be considered in step 208. In some embodiments, only a subset of the possible combinations of optical inspection modes for selected values ​​of N and M are considered in step 208 (e.g., to be provided as input to an unsupervised classifier to generate probabilities for background and defect classes based on associated inspection images). For example, one or more primary modes may be selected such that all combinations of optical inspection modes considered in step 208 include one or more primary modes. It is contemplated herein that selecting one or more primary modes for inclusion in all combinations considered in step 208 can substantially reduce the time required to develop an inspection recipe (e.g., generally, step 202) while also providing inspection performance superior to single-mode inspection, although the precise performance improvement may be affected by the selection of one or more primary modes. The selection of one or more primary modes may be based on any criteria, such as, but not limited to, SNR associated with the defect signal, image contrast, or any image quality metric.

[0080] Any classifier suitable for generating a probability of each location of the preliminary sample being in the background or defect class may be utilized in step 208. For example, the classifier may include any type of machine learning technique, such as, but not limited to, supervised and / or unsupervised machine learning techniques.

[0081] In some embodiments, the classifier includes a soft clustering technique, such as (but not limited to) a Gaussian mixture model (GMM). For example, a GMM can generate a probabilistic model in M-dimensional space and can utilize an expectation-maximization algorithm to iteratively define parameters associated with the Gaussian distribution describing the various classes (or clusters) in an M-dimensional point cloud associated with the M optical detection images in each considered combination. For example, the GMM can assign a negative log-likelihood (NLL) score (2) in Observe at a pixel position Intensity of treatment The expectation maximization algorithm can then adjust the parameters of the probability model for the relevant data to maximize or minimize It should be noted that this technique does not require labeling and can advantageously be implemented in a relatively short timescale and thus retraining (e.g., repeated updating of the detection recipe in step 202) can be achieved in a relatively short timescale (such as, but not limited to, days depending on the hardware used).

[0082] In some embodiments, the classifier includes a supervised discriminant analysis technique, such as, but not limited to, a linear discriminant analysis (LDA) technique. It should be noted that an LDA technique may be appropriate when finding defect signals in a region of M-dimensional space. Furthermore, a supervised technique, including, but not limited to, LDA, may be trained based on inspection images generated using associated optical patterns labeled based on known information, such as known classes of interest (e.g., background or defect class). These labels may correspond to individual pixels and / or groups of pixels corresponding to associated physical locations on the training sample. For example, these labels may optionally describe a region on the training sample having a size corresponding to the point spread function in the associated image.

[0083] Figures 3A to 3E depict performance simulations of one embodiment based on an LDA classifier. However, it should be understood that Figures 3A to 3E and the associated description are provided for illustrative purposes only and should not be construed as limiting.

[0084] FIG3A illustrates a simulated point cloud associated with background and defect probability for a specific defect type identified by an LDA technique using simulated optical inspection images from two different optical modes (e.g., M = 2) generated using a primary mode selected based on SNR, according to one or more embodiments of the present invention. For example, the SNR can (but need not) be determined based on the ratio of the maximum grayscale of an absolute difference image to the standard deviation of the same difference image at the location of an identified defect. In this manner, FIG3A can correspond to a selected combination of optical inspection modes (e.g., from step 210) out of N possible optical inspection modes, where one of the two optical inspection modes is preselected (e.g., as a primary optical inspection mode) based on having the best SNR, and the second is selected (e.g., in step 210) to provide the best separation between classes.

[0085] FIG3B is a histogram of the PDFs of the background and defect classes along line 302 in FIG3A according to one or more embodiments of the present invention. For example, line 302 may be associated with a linear transformation of the point cloud coordinates that best separates the two classes (e.g., a dividing line).

[0086] FIG3C illustrates a simulated point cloud associated with background and defect probabilities for a particular defect type identified using simulated optical inspection images from two different optical modes (e.g., M = 2) generated by an LDA technique using analysis based on all possible combinations of optical modes, according to one or more embodiments of the present invention. For example, FIG3C may correspond to a selected combination of optical inspection modes (e.g., from step 210) among N possible optical inspection modes, where all possible combinations are considered and where this selected combination provides the best separation between background and defect classes.

[0087] FIG3D is a histogram of the PDFs of the background and defect classes along line 304 in FIG3C according to one or more embodiments of the present invention. For example, line 304 may be associated with a linear transformation of the point cloud coordinates that best separates the two classes (e.g., a dividing line).

[0088] FIG3E is a graph of receiver operating characteristics (ROC) for the multi-mode detection configuration of FIG3A-3D and for single-mode operation according to one or more embodiments of the present invention. Specifically, line 306 corresponds to an ROC curve for the multi-mode configuration of FIG3C and FIG3D in which all possible combinations of optical detection modes are evaluated, line 308 corresponds to an ROC curve for the multi-mode configuration of FIG3A and FIG3B in which a primary mode is selected based on SNR and associated combinations are evaluated, and line 310 corresponds to an ROC curve for single-mode detection selected based on SNR.

[0089] An ROC curve is generated by varying a detection threshold used to classify data into different categories, where all events above the threshold become detections. The curve then includes the ratio of true detections (e.g., the area above the threshold in the defect histogram) as the capture rate to the false detection rate (e.g., the area above the threshold in the background histogram) on the horizontal axis. For the single-mode case (line 310), this threshold can be a probability value (e.g., a value of the PDF associated with a defect). For the multi-mode case, this threshold can correspond to a position on a line 312 perpendicular to the associated dividing line.

[0090] As shown in FIG3E , multimode detection using a primary mode selected based on SNR (line 308) improves performance (e.g., reduces false positives for a given capture rate) by several orders of magnitude relative to single-mode detection (line 310), while multimode detection in which all combinations of modes are considered (line 306) further reduces the false positive rate by several additional orders of magnitude.

[0091] Referring again generally to FIG. 2B , in some embodiments, the classifier comprises a neural network (e.g., a generative neural network or the like). It is contemplated herein that a neural network can generally accept any number of input channels and generate relationships between the input data. In this manner, a neural network can output a probability that a particular pixel (and associated location within a sample) belongs to any number of classes (e.g., background or defect). For example, a neural network can generate a PDF and perform an analysis similar to the diagrams in FIG. 3A through FIG. 3E for the point cloud-based soft clustering technique.

[0092] In some embodiments, step 208 includes training or implementing a neural network based on inspection images from various combinations of optical inspection modes. The inspection images may include any combination of original inspection images or difference images associated with each selected combination of optical inspection modes.

[0093] In some embodiments, step 208 further includes training or implementing a neural network based on one or more additional channels of data. For example, the neural network may also receive design data associated with a sample (e.g., a preliminary sample). In this manner, the generated probabilities may be based at least in part on spatial relationships between pixels in the inspection image (e.g., spatial relationships between locations on the sample). As an example, certain defect types may typically occur on or around certain patterned features. In this manner, knowledge of these features may facilitate more accurate differentiation between background and defects at these locations and / or between different types of defects. As another example, certain defect types may typically have an extended size and therefore occupy multiple pixels in an inspection image. In this manner, pixel grouping may be taken into account when generating the probability of any given pixel belonging to the background or defect category. It should be further understood that these examples are merely illustrative and do not limit the operation of a neural network.

[0094] The design data may be in any suitable format, such as, but not limited to, a design clip in the form of an image that provides a representation of a desired feature on the sample. This data may originate from any location, such as, but not limited to, a record or a storage medium (e.g., memory 136). Furthermore, the design data may have any desired level of detail. In some embodiments, the design data provided to a neural network includes relatively low-frequency design information (e.g., design information associated with a spatial scale greater than a selected value or the like). For example, the design data may provide information about the size, shape, and orientation of features on the sample while avoiding the introduction of high-frequency noise into the model.

[0095] FIG4 is a graph of captured defects versus disturbing points (e.g., background) for multi-mode inspection using a neural network (line 402) and single-mode inspection (line 404) according to one or more embodiments of the present invention. As shown in FIG4, for this example, multi-mode inspection using a neural network provides nearly 10 times fewer disturbing points than single-mode inspection.

[0096] Referring again to FIG. 2C , step 204 of testing a test sample using the test recipe developed in step 202 is described in more detail according to one or more embodiments of the present invention.

[0097] In some embodiments, step 204 of inspecting a test sample includes a step 212 of generating M inspection images of the test sample using a selected combination of M optical inspection modes from the inspection recipe (e.g., using one or more optical imaging subsystems 102). Furthermore, the M inspection images of the test sample may be registered and / or scaled so that they have a common number of aligned pixels, where the pixels correspond to locations on the test sample.

[0098] In some embodiments, step 204 of inspecting a test sample includes a step 214 of classifying the test sample's location into background and defect categories based on the M inspection images and the classifier in the inspection recipe. For example, the selected classifier trained in step 208 to generate probabilities for each pixel belonging to background and defect categories can now be used in step 214 to classify defects based on background and defect categories. For example, the selected classifier trained in step 208 can provide weights or equations describing background and defect categories suitable for predicting defects in new inspection images of the test sample.

[0099] In some cases, a classifier trained in step 202 and utilized in step 204 includes a fixed set of trained weights that simply propagate the input image to the output without any feedback. The weights may be applied only to the input data to achieve a linear transformation, and all nonlinear activations are maintained to be the same as those used during training. In other words, the unsupervised classifier may remain in the same state at the end of its training. In some cases, some weights in the trained classifier may be updated to account for variations within a particular test sample or across multiple test samples (e.g., wafer-to-wafer or intra-wafer variations) as well as variations in tool operating conditions.

[0100] In some cases, step 214 performs a soft classification and may generate a probability that each pixel (and therefore each corresponding position of the test sample) belongs to each of the background and defect classes. In some cases, step 214 performs a hard classification and assigns each pixel to a single class (e.g., based on one or more thresholds applied to the probability).

[0101] Furthermore, in applications where the classifier comprises a neural network, step 214 may incorporate additional input channels associated with the test samples, such as, but not limited to, design data.

[0102] Referring again generally to FIG. 2A , it is contemplated herein that the techniques associated with method 200 may be combined with additional unsupervised or supervised techniques for further distinguishing between defects of interest and disturber signals on a test sample. For example, the defects identified on a test sample in step 214 may then be further analyzed using a supervised classifier. In some examples, the defects identified in step 214 may be analyzed using an additional inspection tool (such as, but not limited to, a scanning electron microscope) to verify the classification. This information may then be used as a label when training a supervised classifier.

[0103] The subject matter described herein sometimes depicts various components contained within or connected to other components. It should be understood that these depicted architectures are merely illustrative, and that many other architectures can be implemented to achieve the same functionality. In a conceptual sense, any configuration of components to achieve the same functionality is effectively "associated" so that the desired functionality is achieved. Thus, any two components herein combined to achieve a particular functionality may be considered "associated" with each other so that the desired functionality is achieved, regardless of architecture or intervening components. Similarly, any two components so associated may also be considered "connected" or "coupled" to each other so that the desired functionality is achieved, and any two components capable of being so associated may also be considered "coupleable" to each other so that the desired functionality is achieved. Specific examples of "coupleable" include, but are not limited to, physically interactable and / or physically interacting components, / or wirelessly interactable and / or wirelessly interacting components, and / or logically interactable and / or logically interacting components.

[0104] It is believed that the present invention and many of its attendant advantages will be understood from the foregoing description, and it will be appreciated that various changes may be made in the form, construction, and arrangement of components without departing from the disclosed subject matter or sacrificing all of its material advantages. The forms described are merely illustrative, and the following claims are intended to encompass and include such changes. Furthermore, it should be understood that the invention is defined by the appended claims.

[0105] 100:Detection system 102: Optical Imaging Subsystem 104: Sample 106: Lighting subsystem 108:Lighting 110: Lighting source 112: Photon Collection System 114: Sample Light 116: Detector 118: Lens 120: Optical components 122: Translation stage 124: Lens 126: Optical components 128: Beam Splitter 130:Objective lens 132:Controller 134: Processor 136: Memory 200: Detection method 202: Steps 204: Steps 206: Steps 208: Steps 210: Steps 212: Steps 214: Steps 302: Line 304: Line 306: Line 308: Line 310: Line 312: Line 402: Line 404: Line

Claims

1. A detection system comprising: A controller comprising one or more processors to execute program instructions, thereby causing the one or more processors to: develop a detection recipe using the steps of: receiving N detection images from a preliminary sample associated with one or more optical detection subsystems associated with N different optical detection modes, wherein N is an integer greater than two, wherein each of the N optical detection modes is associated with unique optical imaging parameters of the one or more optical detection subsystems, wherein the position in the N detection images of the preliminary sample corresponds to a position on the preliminary sample; A classifier is used to generate the probability of each location of the preliminary sample in the background or defect category from the detection images of at least some combinations of M optical detection modes, where M is an integer greater than one and less than N and corresponds to the number of optical detection modes to be included in the detection formulation; and one of the combinations of M optical detection modes is selected as a selected combination based on a metric describing the difference between the background and defect categories; and defects on the test sample are identified using M detection images of a test sample generated based on the detection formulation having the selected combination of M optical detection modes.

2. The detection system of claim 1, wherein identifying defects on the test sample using M detection images of the test sample generated based on the detection formula having M optical detection modes of the selected combination includes: Based on the M detection images in the detection formula and the classifier, the test sample is classified into the background and defect categories at the locations therein.

3. The detection system of claim 1, wherein developing the detection formula further includes registering the N detection images of the preliminary sample.

4. The detection system of claim 1, wherein the N detection images of the preliminary sample have a common number of pixels.

5. The detection system of claim 1, wherein the N detection images of the preliminary sample and the M detection images of the test sample correspond to the original images from the one or more optical detection subsystems.

6. The detection system of claim 1, wherein the N detection images of the preliminary sample and the M detection images of the test sample correspond to difference images based on the differences between the original images and reference images from the one or more optical detection subsystems.

7. The detection system as described in Request 1, wherein the defect categories comprise a single defect category.

8. The detection system as described in Request 1, wherein the defect categories comprise two or more defect categories.

9. The detection system as described in claim 1, wherein the classifier comprises: An unsupervised classifier.

10. The detection system of claim 9, wherein the unsupervised classifier implements a soft clustering technique.

11. The detection system as described in claim 10, wherein the soft clustering technology includes: Gaussian Mixture Model (GMM) technique.

12. The detection system as described in claim 1, wherein the classifier comprises: A supervised classifier.

13. The detection system of claim 12, wherein the supervised classifier implements the Linear Discriminant Analysis (LDA) technique.

14. The detection system of claim 12, wherein the supervised classifier is trained on training data containing at least one of labeled background or defect category.

15. The detection system of claim 1, wherein the classifier comprises: A neural network.

16. The detection system of claim 15, wherein using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category from at least some combinations of the M optical detection modes, includes: A neural network is used to generate the probability of each of the locations of the preliminary sample in the background or defect category by using the detection images from at least some combinations of the M optical detection modes.

17. The detection system of claim 1, wherein using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category from at least some combinations of the M optical detection modes, includes: Using this classifier, the detection images from all combinations of the M optical detection modes are used to generate the probability of each of the locations of the preliminary sample in the background or defect category.

18. The detection system of claim 1, wherein developing the detection formula further includes: Selecting one of the optical inspection modes as a primary mode, wherein using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category using inspection images from at least some combinations of the M optical inspection modes includes: using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category using inspection images from at least some combinations of the M optical inspection modes including the primary mode.

19. The detection system of claim 1, wherein the N different optical detection modes are associated with a difference in at least one of an illumination wavelength, an illumination polarization, or an illumination angle.

20. The detection system of claim 1, wherein the N different optical detection modes are associated with a difference in at least one of a wavelength, a polarization, or an angle of light collected by the one or more optical detection subsystems and directed to a detector.

21. The detection system as requested in item 1, where M equals two.

22. The detection system as requested in item 1, wherein N is greater than or equal to three.

23. A detection system comprising: One or more optical detection subsystems; A controller communicatively coupled to the one or more optical inspection subsystems, the controller including one or more processors to execute program instructions, thereby causing the one or more processors to: develop an inspection recipe using the steps of: receiving N inspection images from a preliminary sample of one or more optical inspection subsystems associated with N different optical inspection modes, wherein N is an integer greater than one-two, wherein each of the N optical inspection modes is associated with unique optical imaging parameters of the one or more optical inspection subsystems, wherein the position in the N inspection images of the preliminary sample corresponds to a position on the preliminary sample; A classifier is used to generate the probability of each location of the preliminary sample in the background or defect category from the detection images of at least some combinations of M optical detection modes, where M is an integer greater than one and less than N and corresponds to the number of optical detection modes to be included in the detection formulation; and one of the combinations of M optical detection modes is selected as a selected combination based on a metric describing the difference between the background and defect categories; and defects on the test sample are identified using M detection images of a test sample generated based on the detection formulation having the selected combination of M optical detection modes.

24. The detection system of claim 23, wherein the one or more optical detection subsystems include: Two or more optical detection subsystems.

25. The detection system of claim 23, wherein the one or more optical detection subsystems include: A single optical detection subsystem.

26. The detection system of claim 23, wherein the N detection images of the preliminary sample and the M detection images of the test sample correspond to at least one of original images from one or more optical detection subsystems or difference images based on the differences between the original images and reference images from one or more optical detection subsystems.

27. The detection system of claim 23, wherein the classifier comprises: An unsupervised classifier.

28. The detection system of claim 23, wherein the classifier comprises: A supervised classifier.

29. The detection system of claim 23, wherein the classifier comprises: A neural network.

30. The detection system of claim 23, wherein the N different optical detection modes are associated with a difference in at least one of an illumination wavelength, an illumination polarization, or an illumination angle.

31. The detection system of claim 23, wherein the N different optical detection modes are associated with at least one of a wavelength, a polarization, or an angle of light collected by the one or more optical detection subsystems and directed to a detector.

32. A detection method, comprising: Developing a test formulation using the following steps: generating N test images of a preliminary sample using one or more optical inspection subsystems associated with N different optical inspection modes, where N is an integer greater than two, where each of the N optical inspection modes is associated with unique optical imaging parameters of the one or more optical inspection subsystems, and where the positions in the N test images of the preliminary sample correspond to positions on the preliminary sample; using a classifier to generate the probability of each position of the preliminary sample in the background or defect category using the test images from at least some combinations of M of the optical inspection modes, where M is an integer greater than one and less than N and corresponds to one number of the optical inspection modes to be included in the test formulation; selecting one of the combinations of the M optical inspection modes as a selected combination based on a metric describing the difference between the background and defect categories; and identifying defects on the test sample using M test images of a test sample generated based on the selected combination of the M optical inspection modes.

33. The detection method of claim 32, wherein identifying defects on the test sample using M detection images of the test sample generated based on the detection formula having the selected combination of M optical detection modes includes: Based on the M detection images in the detection formula and the classifier, the test sample is classified into the background and defect categories at the locations therein.

34. The detection method of claim 32, wherein developing the detection formulation further includes registering the N detection images of the preliminary sample.

35. The detection method of claim 32, wherein the N detection images of the preliminary sample have a common number of pixels.

36. The detection method of claim 32, wherein the N detection images of the preliminary sample and the M detection images of the test sample correspond to the original images from the one or more optical detection subsystems.

37. The detection method of claim 32, wherein the N detection images of the preliminary sample and the M detection images of the test sample correspond to difference images based on the differences between the original images and reference images from one or more optical detection subsystems.

38. The detection method as described in claim 32, wherein the defect categories comprise a single defect category.

39. The detection method as described in claim 32, wherein the defect categories comprise two or more defect categories.

40. The detection method as described in request item 32, wherein the classifier comprises: An unsupervised classifier.

41. The detection method as described in request item 32, wherein the classifier comprises: A supervised classifier.

42. The detection method as described in request item 32, wherein the classifier comprises: A neural network.

43. The detection method of claim 32, wherein using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category from at least some combinations of the M optical detection modes of the detection images includes: Using this classifier, the detection images from all combinations of the M optical detection modes are used to generate the probability of each of the locations of the preliminary sample in the background or defect category.

44. The detection method of claim 32, wherein developing the detection formulation further includes: Selecting one of the optical inspection modes as a primary mode, wherein using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category using inspection images from at least some combinations of the M optical inspection modes includes: using the classifier to generate the probability of each of the locations of the preliminary sample in the background or defect category using inspection images from at least some combinations of the M optical inspection modes including the primary mode.

45. The detection method of claim 32, wherein the N different optical detection modes are associated with a difference in at least one of an illumination wavelength, an illumination polarization, or an illumination angle.

46. ​​The detection method of claim 32, wherein the N different optical detection modes are associated with a difference in at least one of a wavelength, a polarization, or an angle of light collected by the one or more optical detection subsystems and directed to a detector.

47. The detection method as described in request item 32, where M equals two.

48. The detection method as requested in item 32, wherein N is greater than or equal to three.

Images