Systems and methods for defect inspection in charged-particle systems

Abstract

Systems and methods for defect inspection including defect detection and defect classification associated with a sample are disclosed. The method for defect inspection includes acquiring a scanned inspection image of a region of a wafer using a charged-particle beam system, and training a deep learning model based on a scanned reference image of the region of the wafer. The trained deep learning model is configured to detect a defect in the acquired scanned inspection image, after detection, extract one or more features of the detected defect, and classify the detected defect based on the one or more extracted features.

Description

SYSTEMS AND METHODS FOR DEFECT INSPECTION IN CHARGED-PARTICLE SYSTEMSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority of US application 63 / 526,457 which was filed on July 13, 2023 and which is incorporated herein in its entirety by reference.TECHNICAL FIELD

[0002] The description herein relates to the field of charged-particle beam systems, and more particularly to systems and methods for detection, inspection, and classification of defects associated with a sample being observed by a charged-particle beam system.BACKGROUND

[0003] In manufacturing processes of integrated circuits (ICs), unfinished or finished circuit components are inspected to ensure that they are manufactured according to design and are substantially free of defects. Inspection systems utilizing optical microscope typically have resolution down to a few hundred nanometers; and the resolution is limited by the wavelength of light. As the physical sizes of IC components continue to reduce down to sub- 100 or even sub- 10 nanometers, inspection systems capable of higher resolution than those utilizing optical microscopes are needed.

[0004] A charged particle (e.g., electron) beam microscope, such as a scanning electron microscope (SEM) or a transmission electron microscope (TEM), capable of resolution down to less than a nanometer, serves as a practicable tool for inspecting IC components having a feature size that is sub- 100 nanometers. With a SEM, electrons of a single primary electron beam, or electrons of a plurality of primary electron beams, can be focused on locations of interest of a semiconductor wafer under inspection. The primary electrons interact with the wafer and may be backscattered or may cause the wafer to emit secondary electrons. The intensity of the electron beams comprising the backscattered electrons and the secondary electrons may vary based on the properties of the internal and external structures of the wafer, and thereby may indicate whether the wafer has defects. To improve defect detection performance, obtaining an accurate SEM image without distortion or misalignment is desired.SUMMARY

[0005] Embodiments of the present disclosure provide apparatuses, systems, and methods for defect detection and classification associated with a sample of charged particle beam systems.

[0006] In some embodiments of the present disclosure, a method for defect inspection is provided. The method includes acquiring a scanned inspection image of a region of a wafer using a charged- particle beam system, and training a deep learning model based on a scanned reference image of theregion of the wafer. The trained deep learning model is configured to detect a defect in the acquired scanned inspection image, after detection, extract one or more features of the detected defect, and classify the detected defect based on the one or more extracted features.

[0007] In some embodiments of the present disclosure, a defect inspection apparatus is provided. The apparatus includes a memory storing a set of instructions and at least one processor configured to execute the set of instructions to cause the apparatus to perform operations. The operations include acquiring a scanned inspection image of a region of a wafer using a charged-particle beam system, and training a deep learning model based on a scanned reference image of the region of the wafer. The trained deep learning model is configured to detect a defect in the acquired scanned inspection image, after detection, extract one or more features of the detected defect, and classify the detected defect based on the one or more extracted features.

[0008] In some embodiments of the present disclosure, a non-transitory computer readable medium that stores a set of instructions that is executable by at least one processor of a computing device to cause the computing device to perform operations for inspecting defects in an inspection image is provided. The operations include acquiring a scanned inspection image of a region of a wafer using a charged-particle beam system, and training a deep learning model based on a scanned reference image of the region of the wafer. The trained deep learning model is configured to detect a defect in the acquired scanned inspection image, after detection, extract one or more features of the detected defect, and classify the detected defect based on the one or more extracted features.

[0009] In some embodiments of the present disclosure, a computer-implemented method for detecting and classifying defects in an image using a deep learning model is provided. The method includes providing a scanned inspection image of a region of a wafer, acquired using a charged- particle beam system, detecting by a trained deep learning model, a defect in the scanned inspection image, extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image, classifying the detected defect based on the extracted one or more features, and generating an output comprising information associated with the detected defect. The trained deep learning model may be trained based on information associated with a defect in a scanned reference image of the region of the wafer, and the at least one characteristic of the scanned reference image may be dissimilar to a corresponding characteristic of the scanned inspection image.

[0010] In some embodiments of the present disclosure, a defect inspection apparatus is provided. The apparatus includes a memory storing a set of instructions and at least one processor configured to execute the set of instructions to cause the apparatus to perform operations. The operations include acquiring a scanned inspection image of a region of a wafer using a charged-particle beam, detecting by a trained deep learning model, a defect in the scanned inspection image, extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image, classifying the detected defect based on the extracted one or more features, and generating an outputcomprising information associated with the detected defect. The trained deep learning model may be trained based on information associated with a defect in a scanned reference image of the region of the wafer, and the at least one characteristic of the scanned reference image may be dissimilar to a corresponding characteristic of the scanned inspection image.

[0011] In some embodiments of the present disclosure, a non-transitory computer readable medium that stores a set of instructions that is executable by at least one processor of a computing device to cause the computing device to perform operations for inspecting defects in an inspection image is provided. The operations include acquiring a scanned inspection image of a region of a wafer using a charged-particle beam, detecting by a trained deep learning model, a defect in the scanned inspection image, extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image, classifying the detected defect based on the extracted one or more features, and generating an output comprising information associated with the detected defect. The trained deep learning model may be trained based on information associated with a defect in a scanned reference image of the region of the wafer, and the at least one characteristic of the scanned reference image may be dissimilar to a corresponding characteristic of the scanned inspection image.

[0012] Other advantages of the embodiments of the present disclosure will become apparent from the following description taken in conjunction with the accompanying drawings wherein are set forth, by way of illustration and example, certain embodiments of the present invention.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Fig. 1 is a schematic diagram illustrating an exemplary electron beam inspection (EBI) system, consistent with embodiments of the present disclosure.

[0014] Fig. 2 is a schematic diagram illustrating an exemplary multi-beam system that is part of the exemplary charged particle beam inspection system of Fig. 1, consistent with embodiments of the present disclosure.

[0015] Fig. 3 illustrates a flowchart representing steps of an exemplary method of image analysis for defect inspection and defect classification.

[0016] Fig. 4 is a schematic diagram of an exemplary system for defect inspection and defect classification, consistent with embodiments of the present disclosure

[0017] Fig. 5 illustrates a flowchart representing steps of an exemplary method of image analysis, consistent with embodiments of the present disclosure.

[0018] Fig. 6 illustrates a flowchart representing steps of an exemplary method of image analysis as shown in Fig. 5 with examples of input and output images, consistent with embodiments of the present disclosure.

[0019] Figs. 7A, 7B, and 7C are exemplary signature maps obtained using different image analysis methods, consistent with embodiments of the present disclosure.

[0020] Fig. 8 is a data graph showing the effectiveness of the proposed method of image analysis, consistent with embodiments of the present disclosure.

[0021] Fig. 9 is a data graph showing the effectiveness of the proposed method of image analysis, consistent with embodiments of the present disclosure.

[0022] Fig. 10 illustrates a process flowchart for an exemplary method of image analysis, consistent with embodiments of the present disclosure.DETAILED DESCRIPTION

[0023] Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings in which the same numbers in different drawings represent the same or similar elements unless otherwise represented. The implementations set forth in the following description of exemplary embodiments do not represent all implementations consistent with the disclosure. Instead, they are merely examples of apparatuses and methods consistent with aspects related to the subject matter recited in the appended claims. For example, although some embodiments are described in the context of utilizing electron beams, the disclosure is not so limited. Other types of charged particle beams may be similarly applied. Furthermore, other imaging systems may be used, such as optical imaging, photodetection, x-ray detection, extreme ultraviolet inspection, deep ultraviolet inspection, or the like.

[0024] Electronic devices are constructed of circuits formed on a piece of silicon called a substrate. Many circuits may be formed together on the same piece of silicon and are called integrated circuits or ICs. The size of these circuits has decreased dramatically so that many more of them can fit on the substrate. For example, an IC chip in a smart phone can be as small as a thumbnail and yet may include over 2 billion transistors, the size of each transistor being less than 1 / 1000th the size of a human hair.

[0025] Making these extremely small ICs is a complex, time-consuming, and expensive process, often involving hundreds of individual steps. Errors in even one step have the potential to result in defects in the finished IC rendering it useless. Thus, one goal of the manufacturing process is to avoid such defects to maximize the number of functional ICs made in the process, that is, to improve the overall yield of the process.

[0026] One component of improving yield is monitoring the chip making process to ensure that it is producing a sufficient number of functional integrated circuits. One way to monitor the process is to inspect the chip circuit structures at various stages of their formation. Inspection may be carried out using a scanning electron microscope (SEM). A SEM can be used to image these extremely small structures, in effect, taking a “picture” of the structures of the wafer. The image can be used to determine if the structure was formed properly and also if it was formed at the proper location. If the structure is defective, then the process can be adjusted so the defect is less likely to recur. Defectsmay be generated during various stages of semiconductor processing. For the reason stated above, it is important to find defects accurately, efficiently, and as early as possible.

[0027] The working principle of a SEM is similar to a camera. A camera takes a picture by receiving and recording brightness and colors of light reflected or emitted from people or objects. A SEM takes a “picture” by receiving and recording energies or quantities of electrons reflected or emitted from the structures. Before taking such a “picture,” an electron beam may be provided onto the structures, and when the electrons are reflected or emitted (“exiting”) from the structures, a detector of the SEM may receive and record the energies or quantities of those electrons to generate an image. To take such a “picture,” some SEMs use a single electron beam (referred to as a “single-beam SEM”), while some SEMs use multiple electron beams (referred to as a “multi-beam SEM”) to take multiple “pictures” of the wafer. By using multiple electron beams, the SEM may provide more electron beams onto the structures for obtaining these multiple “pictures,” resulting in more electrons exiting from the structures. Accordingly, the detector may receive more exiting electrons simultaneously, and generate images of the structures of the wafer with a higher efficiency and a faster speed.

[0028] For example, voltage contrast inspection may be used as an early proxy for electric yield associated with a sample. SEM images including voltage contrast patterns typically show a random occurrence of failures associated with features of a sample (e.g., varying grey scale levels of features). For example, grey level intensity levels in an SEM inspection image may deviate from grey level intensity levels in a defect-free SEM image, thereby indicating that a sample associated with the SEM inspection image includes one or more defects (e.g., electrical open or short failures). In some embodiments, other characteristics (e.g., besides or in addition to voltage contrast characteristics) in an SEM inspection image may deviate from a defect-free SEM image (e.g., characteristics related to line-edge roughness, line-width roughness, local critical dimension uniformity, necking, bridging, edge placement errors, etc.), thereby indicating that a sample associated with the SEM inspection image includes one or more defects.

[0029] The existing defect inspection and classification methods suffer from several drawbacks. First, the existing defect inspection processes rely on human intervention, rendering the process inconsistent and time-consuming and also negatively impacting the inspection throughput. For example, in both die-to-die (D2D) and die-to-database (D2DB) techniques, in the alignment stage (e.g., step 320 of process 300), the SEM image of the wafer being inspected has to be accurately aligned with the corresponding reference images before any detection can be performed. In some cases, the images have to be re-aligned at later stages as well. Further, during the defect classification stages, human effort is relied upon to build effective feature extraction methods, to choose a set of representative training data, and to build effective classifiers. Second, the existing defect inspection pipeline is complicated and long, rendering it susceptible to high nuisance rates or low capture rates. The inspection SEM images, and the reference images may have large differences in gray-level,brightness, pattern fidelity, pixel size, resolution, sharpness, among other characteristics. One or more of these differences in image characteristics may mislead the deep learning algorithms to generate a large number of false positives, resulting in low capture rates. Third, the existing techniques are not self-adaptive to various pixel sizes, resolutions, and imaging conditions. In addition, training multiple models which can be applied to images obtained under different imaging conditions or having different characteristics may be extremely expensive and impractical. Therefore, it may be desirable to provide systems and methods for image analysis to detect and classify defects in an image using a deep learning framework which addresses one or more of the above-mentioned drawbacks of the existing techniques.

[0030] Some embodiments of the present disclosure are directed to a method of image analysis using a trained deep learning model to jointly detect and classify defects in an inspection image. The method includes acquiring an inspection SEM image and training a deep learning model to detect a defect in the inspection image, extracting one or more features of the detected defect, and classifying the defect based on the extracted features. The proposed method, unlike existing defect detection and inspection methods, uses a single-stage deep learning model configured to jointly detect and classify defects in an image. The single-stage deep learning model may be trained, using data augmentation techniques, for example, such that it can be effectively applied to inspection images acquired under different imaging conditions even when the training data comprises a single acquisition condition. One or more steps of the proposed method for image analysis is implemented by an end-to-end deep learning model that is a generic, self-adaptive, light-weight, and nuisance aware, configured to enhance wafer inspection throughput, efficiency, and accuracy.

[0031] Relative dimensions of components in drawings may be exaggerated for clarity. Within the following description of drawings, the same or like reference numbers refer to the same or like components or entities, and only the differences with respect to the individual embodiments are described. As used herein, unless specifically stated otherwise, the term “or” encompasses all possible combinations, except where infeasible. For example, if it is stated that a component may include A or B, then, unless specifically stated otherwise or infeasible, the component may include A, or B, or A and B. As a second example, if it is stated that a component may include A, B, or C, then, unless specifically stated otherwise or infeasible, the component may include A, or B, or C, or A and B, or A and C, or B and C, or A and B and C.

[0032] Fig. 1 illustrates an exemplary electron beam inspection (EBI) system 100 consistent with embodiments of the present disclosure. EBI system 100 may be used for imaging. As shown in Fig. 1, EBI system 100 includes a main chamber 101, a load / lock chamber 102, an electron beam tool 104, and an equipment front end module (EFEM) 106. Electron beam tool 104 is located within main chamber 101. EFEM 106 includes a first loading port 106a and a second loading port 106b. EFEM 106 may include additional loading port(s). First loading port 106a and second loading port 106breceive wafer front opening unified pods (FOUPs) that contain wafers (e.g., semiconductor wafers or wafers made of other material(s)) or samples to be inspected (wafers and samples may be used interchangeably). A “lot” is a plurality of wafers that may be loaded for processing as a batch.

[0033] One or more robotic arms (not shown) in EFEM 106 may transport the wafers to load / lock chamber 102. Load / lock chamber 102 is connected to a load / lock vacuum pump system (not shown) which removes gas molecules in load / lock chamber 102 to reach a first pressure below the atmospheric pressure. After reaching the first pressure, one or more robotic arms (not shown) may transport the wafer from load / lock chamber 102 to main chamber 101. Main chamber 101 is connected to a main chamber vacuum pump system (not shown) which removes gas molecules in main chamber 101 to reach a second pressure below the first pressure. After reaching the second pressure, the wafer is subject to inspection by electron beam tool 104. Electron beam tool 104 may be a single-beam system or a multi-beam system.

[0034] A controller 109 is electronically connected to electron beam tool 104. Controller 109 may be a computer configured to execute various controls of EBI system 100. While controller 109 is shown in Fig. 1 as being outside of the structure that includes main chamber 101, load / lock chamber 102, and EFEM 106, it is appreciated that controller 109 may be a part of the structure.

[0035] In some embodiments, controller 109 may include one or more processors (not shown). A processor may be a generic or specific electronic device capable of manipulating or processing information. For example, the processor may include any combination of any number of a central processing unit (or “CPU”), a graphics processing unit (or “GPU”), an optical processor, a programmable logic controllers, a microcontroller, a microprocessor, a digital signal processor, an intellectual property (IP) core, a Programmable Logic Array (PLA), a Programmable Array Logic (PAL), a Generic Array Logic (GAL), a Complex Programmable Logic Device (CPLD), a Field- Programmable Gate Array (FPGA), a System On Chip (SoC), an Application-Specific Integrated Circuit (ASIC), and any type circuit capable of data processing. The processor may also be a virtual processor that includes one or more processors distributed across multiple machines or devices coupled via a network.

[0036] In some embodiments, controller 109 may further include one or more memories (not shown). A memory may be a generic or specific electronic device capable of storing codes and data accessible by the processor (e.g., via a bus). For example, the memory may include any combination of any number of a random-access memory (RAM), a read-only memory (ROM), an optical disc, a magnetic disk, a hard drive, a solid-state drive, a flash drive, a security digital (SD) card, a memory stick, a compact flash (CF) card, or any type of storage device. The codes may include an operating system (OS) and one or more application programs (or “apps”) for specific tasks. The memory may also be a virtual memory that includes one or more memories distributed across multiple machines or devices coupled via a network.

[0037] Reference is now made to Fig. 2, which is a schematic diagram illustrating an exemplary electron beam tool 104 including a multi-beam inspection tool that is part of the EBI system 100 of Fig. 1, consistent with embodiments of the present disclosure. In some embodiments, electron beam tool 104 may be operated as a single-beam inspection tool that is part of EBI system 100 of Fig. 1. Multi-beam electron beam tool 104 (also referred to herein as apparatus 104) comprises an electron source 201, a Coulomb aperture plate (or “gun aperture plate”) 271, a condenser lens 210, a source conversion unit 220, a primary projection system 230, a motorized stage 209, and a sample holder 207 supported by motorized stage 209 to hold a sample 208 (e.g., a wafer or a photomask) to be inspected. Multi-beam electron beam tool 104 may further comprise a secondary projection system 250 and an electron detection device 240. Primary projection system 230 may comprise an objective lens 231. Electron detection device 240 may comprise a plurality of detection elements 241, 242, and 243. A beam separator 233 and a deflection scanning unit 232 may be positioned inside primary projection system 230.

[0038] Electron source 201, Coulomb aperture plate 271, condenser lens 210, source conversion unit 220, beam separator 233, deflection scanning unit 232, and primary projection system 230 may be aligned with a primary optical axis 204 of apparatus 104. Secondary projection system 250 and electron detection device 240 may be aligned with a secondary optical axis 251 of apparatus 104.

[0039] Electron source 201 may comprise a cathode (not shown) and an extractor or anode (not shown), in which, during operation, electron source 201 is configured to emit primary electrons from the cathode and the primary electrons are extracted or accelerated by the extractor and / or the anode to form a primary electron beam 202 that form a primary beam crossover (virtual or real) 203. Primary electron beam 202 may be visualized as being emitted from primary beam crossover 203.

[0040] Source conversion unit 220 may comprise an image-forming element array (not shown), an aberration compensator array (not shown), a beam-limit aperture array (not shown), and a pre-bending micro-deflector array (not shown). In some embodiments, the pre-bending micro-deflector array deflects a plurality of primary beamlets 211, 212, 213 of primary electron beam 202 to normally enter the beam-limit aperture array, the image-forming element array, and an aberration compensator array. In some embodiments, apparatus 104 may be operated as a single -beam system such that a single primary beamlet is generated. In some embodiments, condenser lens 210 is designed to focus primary electron beam 202 to become a parallel beam and be normally incident onto source conversion unit 220. The image-forming element array may comprise a plurality of micro-deflectors or micro-lenses to influence the plurality of primary beamlets 211, 212, 213 of primary electron beam 202 and to form a plurality of parallel images (virtual or real) of primary beam crossover 203, one for each of the primary beamlets 211, 212, and 213. In some embodiments, the aberration compensator array may comprise a field curvature compensator array (not shown) and an astigmatism compensator array (not shown). The field curvature compensator array may comprise a plurality of micro-lenses tocompensate field curvature aberrations of the primary beamlets 211, 212, and 213. The astigmatism compensator array may comprise a plurality of micro-stigmators to compensate astigmatism aberrations of the primary beamlets 211, 212, and 213. The beam-limit aperture array may be configured to limit diameters of individual primary beamlets 211, 212, and 213. Fig. 2 shows three primary beamlets 211, 212, and 213 as an example, and it is appreciated that source conversion unit 220 may be configured to form any number of primary beamlets. Controller 109 may be connected to various parts of EBI system 100 of Fig. 1, such as source conversion unit 220, electron detection device 240, primary projection system 230, or motorized stage 209. In some embodiments, as explained in further details below, controller 109 may perform various image and signal processing functions. Controller 109 may also generate various control signals to govern operations of the charged particle beam inspection system.

[0041] Condenser lens 210 is configured to focus primary electron beam 202. Condenser lens 210 may further be configured to adjust electric currents of primary beamlets 211, 212, and 213 downstream of source conversion unit 220 by varying the focusing power of condenser lens 210. Alternatively, the electric currents may be changed by altering the radial sizes of beam-limit apertures within the beam-limit aperture array corresponding to the individual primary beamlets. The electric currents may be changed by both altering the radial sizes of beam-limit apertures and the focusing power of condenser lens 210. Condenser lens 210 may be an adjustable condenser lens that may be configured so that the position of its first principal plane is movable. The adjustable condenser lens may be configured to be magnetic, which may result in off-axis beamlets 212 and 213 illuminating source conversion unit 220 with rotation angles. The rotation angles change with the focusing power or the position of the first principal plane of the adjustable condenser lens. Condenser lens 210 may be an anti-rotation condenser lens that may be configured to keep the rotation angles unchanged while the focusing power of condenser lens 210 is changed. In some embodiments, condenser lens 210 may be an adjustable anti-rotation condenser lens, in which the rotation angles do not change when its focusing power and the position of its first principal plane are varied.

[0042] Objective lens 231 may be configured to focus beamlets 211, 212, and 213 onto a sample 208 for inspection and may form, in the current embodiments, three probe spots 221, 222, and 223 on the surface of sample 208. Coulomb aperture plate 271, in operation, is configured to block off peripheral electrons of primary electron beam 202 to reduce Coulomb effect. The Coulomb effect may enlarge the size of each of probe spots 221, 222, and 223 of primary beamlets 211, 212, 213, and therefore deteriorate inspection resolution.

[0043] Beam separator 233 may, for example, be a Wien filter comprising an electrostatic deflector generating an electrostatic dipole field and a magnetic dipole field (not shown in Fig. 2). In operation, beam separator 233 may be configured to exert an electrostatic force by electrostatic dipole field on individual electrons of primary beamlets 211, 212, and 213. The electrostatic force is equal inmagnitude but opposite in direction to the magnetic force exerted by magnetic dipole field of beam separator 233 on the individual electrons. Primary beamlets 211, 212, and 213 may therefore pass at least substantially straight through beam separator 233 with at least substantially zero deflection angles.

[0044] Deflection scanning unit 232, in operation, is configured to deflect primary beamlets 211, 212, and 213 to scan probe spots 221, 222, and 223 across individual scanning areas in a section of the surface of sample 208. In response to incidence of primary beamlets 211, 212, and 213 or probe spots 221, 222, and 223 on sample 208, electrons emerge from sample 208 and generate three secondary electron beams 261, 262, and 263. Each of secondary electron beams 261, 262, and 263 typically comprise secondary electrons (having electron energy < 50eV) and backscattered electrons (having electron energy between 50eV and the landing energy of primary beamlets 211, 212, and 213). Beam separator 233 is configured to deflect secondary electron beams 261, 262, and 263 towards secondary projection system 250. Secondary projection system 250 subsequently focuses secondary electron beams 261, 262, and 263 onto detection elements 241, 242, and 243 of electron detection device 240. Detection elements 241, 242, and 243 are arranged to detect corresponding secondary electron beams 261, 262, and 263 and generate corresponding signals which are sent to controller 109 or a signal processing system (not shown), e.g., to construct images of the corresponding scanned areas of sample 208.

[0045] In some embodiments, detection elements 241, 242, and 243 detect corresponding secondary electron beams 261, 262, and 263, respectively, and generate corresponding intensity signal outputs (not shown) to an image processing system (e.g., controller 109). In some embodiments, each detection element, 241, 242, and 243 may comprise one or more pixels. The intensity signal output of a detection element may be a sum of signals generated by all the pixels within the detection element.

[0046] In some embodiments, controller 109 may comprise image processing system that includes an image acquirer (not shown), a storage (not shown). The image acquirer may comprise one or more processors. For example, the image acquirer may comprise a computer, a server, a mainframe host, terminals, a personal computer, any kind of mobile computing devices, and the like, or a combination thereof. The image acquirer may be communicatively coupled to electron detection device 240 of apparatus 104 through a medium such as an electrical conductor, an optical fiber cable, a portable storage media, IR, Bluetooth, internet, a wireless network, a wireless radio, among others, or a combination thereof. In some embodiments, the image acquirer may receive a signal from electron detection device 240 and may construct an image. The image acquirer may thus acquire images of sample 208. The image acquirer may also perform various post-processing functions, such as generating contours, superimposing indicators on an acquired image, and the like. The image acquirer may be configured to perform adjustments of brightness and contrast, etc. of acquired images. In some embodiments, the storage may be a storage medium such as a hard disk, flash drive, cloudstorage, random access memory (RAM), other types of computer readable memory, and the like. The storage may be coupled with the image acquirer and may be used for saving scanned raw image data as original images, and post-processed images.

[0047] In some embodiments, the image acquirer may acquire one or more images of a sample based on an imaging signal received from electron detection device 240. An imaging signal may correspond to a scanning operation for conducting charged particle imaging. An acquired image may be a single image comprising a plurality of imaging areas. The single image may be stored in the storage. The single image may be an original image that may be divided into a plurality of regions. Each of the regions may comprise one imaging area containing a feature of sample 208. The acquired images may comprise multiple images of a single imaging area of sample 208 sampled multiple times over a time sequence. The multiple images may be stored in the storage. In some embodiments, controller 109 may be configured to perform image processing steps with the multiple images of the same location of sample 208.

[0048] In some embodiments, controller 109 may include measurement circuitries (e.g., analog-to- digital converters) to obtain a distribution of the detected secondary electrons. The electron distribution data collected during a detection time window, in combination with corresponding scan path data of each of primary beamlets 211, 212, and 213 incident on the wafer surface, can be used to reconstruct images of the wafer structures under inspection. The reconstructed images can be used to reveal various features of the internal or external structures of sample 208, and thereby can be used to reveal any defects that may exist in the wafer.

[0049] In some embodiments, controller 109 may control motorized stage 209 to move sample 208 during inspection of sample 208. In some embodiments, controller 109 may enable motorized stage 209 to move sample 208 in a direction continuously at a constant speed. In other embodiments, controller 109 may enable motorized stage 209 to change the speed of the movement of sample 208 overtime depending on the steps of scanning process.

[0050] Although Fig. 2 shows that apparatus 104 uses three primary electron beams, it is appreciated that apparatus 104 may use one or more number of primary electron beams. The present disclosure does not limit the number of primary electron beams used in apparatus 104. In some embodiments, apparatus 104 may be a SEM used for lithography, defect inspection, or a combination thereof.

[0051] Compared with a single charged-particle beam imaging system (“single-beam system”), a multiple charged-particle beam imaging system (“multi-beam system”) may be designed to optimize throughput for different scan modes. Embodiments of this disclosure provide a multi-beam system with the capability of optimizing throughput for different scan modes by using beam arrays with different geometries, adapting to different throughputs and resolution requirements. It is to be appreciated that a single -beam system may be used as well to acquire one or more images of asubstrate or a region of interest on a semiconductor wafer during fabrication of devices, as appropriate.

[0052] A non-transitory computer readable medium may be provided that stores instructions for a processor (e.g., processor of controller 109 of Figs. 1-2) to carry out image processing, data processing, beamlet scanning, database management, graphical display, operations of a charged particle beam apparatus, or another imaging device, or the like. Common forms of non-transitory media include, for example, a floppy disk, a flexible disk, hard disk, solid state drive, magnetic tape, or any other magnetic data storage medium, a CD-ROM, any other optical data storage medium, any physical medium with patterns of holes, a RAM, a PROM, and EPROM, a FLASH-EPROM or any other flash memory, NVRAM, a cache, a register, any other memory chip or cartridge, and networked versions of the same.

[0053] Reference is now made to Fig. 3, which illustrates a flowchart representing an exemplary image analysis method. An image analysis process 300, as shown in Fig. 3, is often used to detect one or more defects and identify the location of the detected defects on a sample.

[0054] At step 310, an inspection system (e.g., EBI system 100 of Fig. 1) may obtain an inspection image 302 (e.g., a SEM image generated during inspection of a sample) and a reference image 304. For example, a reference image may be a defect-free SEM image of a sample. A reference image may include one or more regions of a sample in a field-of-view (FOV). Reference image 304 may include one or more manually-labeled high-resolution reference SEM images, or images of a reference die, or the like. For example, a user may capture multiple high-resolution SEM images of a region of a sample using an inspection system and manually label a position of a feature based on the mask information or information from a Graphic Database System (GDS) file containing information associated with a layout of the devices on the wafer, for example. In some embodiments, reference image 304 may be a layout file for a wafer design corresponding to the inspection image. The layout file may be a golden image or in a Graphic Database System (GDS) format, Graphic Database System II (GDS II) format, an Open Artwork System Interchange Standard (OASIS) format, or a Caltech Intermediate Format (CIF). The sample design may include patterns or structures for inclusion on the sample. In some embodiments, a layout in GDS or OASIS format, among other formats, may comprise feature information stored in a binary file format representing planar geometric shapes, text, and other information related to the sample design.

[0055] At step 320, inspection image 302 is aligned with a labeled reference image 304 including at least one or more features of inspection image. In some cases, a processor of inspection system may perform the alignment of images to identify the locations of one or more defects on a sample being inspected.

[0056] At step 330, inspection system detects one or more defects on an inspection sample by comparing the aligned images to a plurality of reference images (e.g., comparing an inspection imageto two defect-free images of a sample during die-to-die inspection). An inspection algorithm may be configured to capture the anomaly or defects in the inspection image 302 based on a comparison with one or more template or reference images. In some cases, the inspection system may additionally perform distortion correction on inspection image 302. Distortion in inspection image 302 may occur because of several reasons including, but not limited to, system operating conditions, tooling factors, calibrations, sample processing history, among other factors.

[0057] At step 340, the inspection system may generate a list of defects based on the comparison performed in step 330. After the defects are detected and listed, a defect classification step may be performed to extract the defect features and identify or classify the detected defects using an automated defect classification (ADC) or semi-automated defect classification technique.

[0058] At step 350, a feature extraction method is performed on the detected defects from inspection image 302. In some cases, the feature extraction is performed manually by human effort. Alternatively, features may be extracted by choosing a set of representative training data, or by choosing one or more classifiers.

[0059] At step 360, the inspection system, using automatic defect classification (ADC), classifies the detected defects based on the features extracted and prior ground truth information and generates a list of predicted defect types.

[0060] There may be several challenges in detecting and classifying defects using a manually labeled reference image 304 in process 300, such as generating a representative reference image, misalignment of the inspection image and the reference image, etc. Generating a labeled representative reference image may include several steps such as, but not limited to, collecting multiple high-resolution SEM images of a region of interest, drawing mask information associated with the region of interest, counting column numbers and row numbers, and labeling features accordingly. One or more of these steps are performed manually by a user or a group of users, making the process inefficient, cumbersome, and error-prone. In some instances, the region of interest may not be covered by a single SEM reference image and one or more reference images may be “stitched” or combined to adequately represent the region of interest. This may make the process even more inefficient and inconsistent. Further, once imaged, the inspection area, scan width, scan rate, inspection modes, etc. of the captured reference SEM images cannot be changed. Furthermore, one or more reference SEM images may suffer from drift and distortion aberrations caused partly by surfacecharging, which can severely impact spatial resolution and critical dimension (CD) measurements. Although digital image correction techniques may be employed to address the drift and distortion artifacts, such techniques are time-consuming and may further introduce variability and negatively impact inspection throughput.

[0061] Fig. 4 is a schematic diagram of an exemplary system 400 for defect detection and classification, consistent with embodiments of the present disclosure. System 400 may include, amongother components, an inspection system 410, a single-stage detector 405 including a detection component 420, and defect classification component 440. Two or more of inspection system 410, detection component 420, and defect classification component 440 may be electrically coupled (directly or indirectly) to each other, either physically (e.g., by a cable) or remotely (e.g., through a wireless connection) such that information may be transmitted or shared between individual components. In some embodiments, inspection system 410 may be the system described with respect to Figs. 1 and 2, used to acquire images of a wafer (see, e.g., sample 208 of Fig. 2). In some embodiments, inspection system 410 may include an image acquisition system analogous to an image acquirer (e.g., in controller 109 of Fig. 1).

[0062] In some embodiments, EBI system 100 may be a SEM system configured to acquire SEM images of a wafer, a region of a wafer, or a region of interest of the wafer. SEM system may be configured to obtain inspection images of regions of wafer for defect inspection or reference images of regions of a wafer to train a deep learning model that may be configured to modify the reference images, if needed. As used herein, a SEM inspection image may be referred to as a scanned inspection image, acquired by scanning or rastering the electron beam across a region of the wafer. In some embodiments, a SEM reference image used to train the deep learning model may be referred to as a scanned reference image, acquired by scanning a region of the wafer. In some embodiments, one or more characteristics of a scanned inspection image and a scanned reference image be different. For example, a scanned inspection image may comprise a low-resolution image acquired by scanning the electron beam across a region of the wafer fewer times than a scanned reference image which may be a high-resolution image. In some embodiments, SEM system may be configured to enable a first scan for acquiring one or more scanned inspection images and to enable a second scan for acquiring one or more scanned reference images. The first scan may include a low number of scans and the second scan may include a higher number of scans, resulting in a better quality of the image acquired by the second scan. The better quality may be a result of higher resolution, higher contrast, higher sensitivity, higher brightness, lower noise-level, larger secondary electron signal, etc. or a combination of these factors. The low number of scans may be a number of scans in a range of about, e.g., 1 to about 10 scans.

[0063] In some embodiments, components of system 400 may be implemented as one or more servers, where each server includes its own processor. In some embodiments, components of system 400 may be implemented as software that may obtain data from one or more databases of system 400. In some embodiments, system 400 may include a single server or a plurality of servers. In some embodiments, system 400 may include one or more modules that are implemented by a controller (e.g., controller 109 of Fig. 1, controller 109 of Fig. 2).

[0064] Inspection system 410 may be configured to transmit data associated with an acquired image, or transmit acquired inspection images of a sample (e.g., sample 208 of Fig. 2), or both, to one or more components of system 400.

[0065] System 400 may further include a defect detection component 420 in communication with inspection system 410 and defect classification component 440. Defect detection component 420 may include one or more processors (referred to as processor 422) and a storage 424. Defect detection component 420 may also include a communication interface 426 to send data to or receive data from inspection system 410 or defect classification component 430. In some embodiments, processor 422 may be configured to perform one or more functions including, but not limited to, identifying one or more features of inspection image, detecting a defect in an inspection image, transmitting information associated with a defect or an inspection image to a deep learning model, among other things. Processor 422 may include any combination of any number of a central processing unit (or “CPU”), a graphics processing unit (or “GPU”), a neural-network processing unit (“NPU”) an optical processor, a programmable logic controllers, a microcontroller, a microprocessor, a digital signal processor, an intellectual property (IP) core, a Programmable Logic Array (PLA), a Programmable Array Logic (PAL), a Generic Array Logic (GAL), a Complex Programmable Logic Device (CPLD), a Field- Programmable Gate Array (FPGA), a System On Chip (SoC), an Application-Specific Integrated Circuit (ASIC), and any type circuit capable of data processing.

[0066] A deep learning model may be trained to perform defect detection by an object detection, or an object recognition technique. In some embodiments, processor 422 of defect detection component 420 may be configured to implement object detection algorithms in a deep learning framework The backend network may be any suitable deep learning network such as, but not limited to, regions with Convolutional Neural Networks (R-CNN), You Only Look Once (YOLO), Fast R-CNN, Faster R- CNN, Mask R-CNN, Spatial Pyramid Pooling (SPP-net), etc., capable of performing object detection and recognition based on use cases.

[0067] Object detection is the process of finding and classifying objects in an image. Object detection using deep learning models may be performed using single-stage or double-stage detection algorithms. Convolutional neural networks (CNN), and specifically, R-CNN approach, is a two-stage detection algorithm in which rectangular region proposals (i.e., regions in the image that may contain an object) with convolutional neural network features are combined. The first stage is configured to identify a subset of regions in an image that may contain an object. The second stage classifies the object in each region. Generally, deep learning models for object detection using regions with CNN are based on three processes: (a) finding regions in the image that may contain an object; (b) extracting CNN features from the regions; and (c) classifying the objects using the extracted features. However, in a single-stage detection algorithm such as YOLO, for example, these three processes are performed in a single operation. As an example, an input image may be fed to the deep learningnetwork and an output of the class probabilities and bounding box coordinates may be generated from the entire image in a single pass without applying a region-proposal network on the features. Advantageously, single-stage object detection deep learning algorithms may predict fewer false positives (i.e., lower nuisance rate) in the background compared to double-stage object detection algorithms which include a separate step for classification of the identified object. As a result, such models are light-weight and have shorter inference times in comparison with other deep learning networks.

[0068] Some embodiments of this disclosure are directed to a deep learning network for single-stage and reference-free defect detection and classification. Further, some embodiments of this disclosure are directed to a deep learning network for single stage, reference-free, detection and classification of defects in images acquired at different conditions and having different characteristics compared to the training data for the deep learning model.

[0069] Defect classification component 440 may include one or more processors (referred to as processor 442) and a storage 444. Storage 444 may be configured to store the data associated with the detected defect or classified defects. Processor 442 may be configured to receive information from or transmit information to detection component 420 or inspection system 410. Additionally, or alternatively, processor 442 may be configured to classify the detected defects and generate a list of defects and associated information including, but not limited to, a location, a size, a type, or a confidence score of the detected defect(s). Processor 442 may include any combination of any number of a central processing unit (or “CPU”), a graphics processing unit (or “GPU”), a neural-network processing unit (“NPU”) an optical processor, a programmable logic controllers, a microcontroller, a microprocessor, a digital signal processor, an intellectual property (IP) core, a Programmable Logic Array (PLA), a Programmable Array Logic (PAL), a Generic Array Logic (GAL), a Complex Programmable Logic Device (CPLD), a Field-Programmable Gate Array (FPGA), a System On Chip (SoC), an Application-Specific Integrated Circuit (ASIC), and any type circuit capable of data processing.

[0070] Reference is now made to Fig. 5, which illustrates a flowchart representing steps of an exemplary method 500 of image analysis, consistent with embodiments of the present disclosure. In some embodiments, one or more steps of image analysis method 500, also referred to herein as method 500, may be performed by EBI system 100 of Fig. 1, a processor associated with EBI system 100, or a defect detection and classification system (e.g., system 400 of Fig. 4). One or more steps, not illustrated in Fig. 5, may be added, deleted, edited, or ordered differently, as appropriate.

[0071] At step 510, an input image 502 may be obtained from an inspection system or an image acquisition system (e.g., image acquirer of controller 109 of EBI system 100 of Fig. 1). In some embodiments, input image 502 may be an image to be inspected or an inspection image (analogous to inspection image 302 of Fig. 3). Input image 502 may comprise a charged-particle beam imageacquired using a charged-particle beam system such as, for example, a SEM system. In some embodiments, input image 502 may be a SEM image, or a backscattered electron (BSE) image, or an image formed using charged-particles originating from a region of a sample upon irradiation with a primary charged-particle beam or a probing beam. In some embodiments, input image 502 may be acquired under an acquisition condition of a charged-particle beam system, for example, a SEM. The acquisition condition may comprise a set of operating parameters including, but not limited to, probe beam current, brightness, contrast, accelerating voltage, lens focus, probe -beam spot size, among other things. The acquisition condition under which input image 502 may be acquired may cause input image 502 to have a characteristic such as a gray level, a brightness, a contrast, a pixel size, a resolution, a signal-to-noise ratio (SNR), a sharpness, etc. It is to be appreciated that in the context of this disclosure, acquisition condition and imaging condition may refer to the conditions or operating parameters of the image acquirer when the image is captured and may be interchangeably used.

[0072] In some embodiments, input image 502 may be a low-resolution SEM image or a quick scan image obtained to perform a high-speed inspection. In some embodiments, input image 502 may be a high-resolution SEM image or a slow scan image obtained to perform a thorough image inspection for a detailed image analysis. In some embodiments, image analysis may include defect inspection. Defect, as used herein, may include bridging, necking, discontinuity, broken lines or patterns, line edge roughness, edge placement errors, or any anomaly that affects the functionality of devices built during the fabrication process.

[0073] Step 510 may further include training a deep learning model to detect and classify defects in an inspection image such as input image 502. In some embodiments, a deep learning model may include an artificial neural network, or a single-stage deep learning framework configured to jointly detect and classify defects in input image 502. Training a deep learning model may include providing one or more reference images to the deep learning model. In some embodiments, reference images may comprise high-resolution SEM images having a small average pixel size, or low-resolution SEM images having a large average pixel size, a low-brightness SEM image, a high-brightness SEM image, a low-contrast SEM image, a high-contrast SEM image, a low SNR SEM image, a high SNR SEM image, or a combination thereof. Reference image may be a SEM image of a region of a wafer, or a region of interest on the wafer (e.g., a semiconductor substrate). In some embodiments, reference images may be simulated images.

[0074] In some embodiments, a reference image may include one or more defects. Training a deep learning model may include providing information associated with the one or more defects in the reference image including, but not limited to, a defect location, a defect size, a defect type, among other things, to the deep learning model. Information associated with defect location may include two- dimensional coordinates (x-y coordinates) of a die containing the defect, x-y coordinates of the location of the defect on the wafer, x-y coordinates with respect to an alignment mark or a referencemark on the wafer, or the like. In some embodiments, information associated with defect size may include dimensions of the defect, shape of the defect, size of the defect with respect to the number of devices affected, among other things. Information associated with the type of defects may include a bridge, a neck, a broken pattern or a broken line, out-of-tolerance line edge roughness, out-of - tolerance edge placement error, etc.

[0075] The deep learning model may be “trained” when the prediction of defect location, defect size, or defect type is substantially convergent with ground truth information. In the context of this disclosure, the term “substantially convergent” refers to a negligible difference between the ground truth information and the predicted information. For example, if the ground truth information associated with a size of the defect (e.g., a break in the line pattern) is 2.5 nm, a predicted defect size of 2.5+0.1 nm would be negligibly different such that the model is substantially convergent with the ground truth information. In this context, ground truth information refers to the target for training or validating a model with a labeled dataset. Developing ground truth datasets may include model design, data labeling, classifier design, and training. Ground truth labels for datasets may be manually annotated by humans (e.g., human reviewed), or an authorized inspection engineer. Ground truth information is provided by direction observation, measurement, and empirical evidence.

[0076] In some embodiments, training a deep learning model may further include performing one or more data augmentation techniques such as, but not limited to, noise-adding, pixel sizing, downsampling, up-sampling, image flipping, image scaling, image cropping, image rotation, adjusting contrast, adjusting brightness, adjusting saturation, performing translation, among other techniques, to reference images. Applying data augmentation techniques to reference images may allow creation of a larger dataset of reference images from a small set of captured images. In some embodiments, data augmentation techniques may be applied to reference images having a characteristic (by virtue of being captured under a condition) to create one or more reference images having a different characteristic. By doing so, the model may be trained with a large variety of conditions without acquiring an image for every condition the model is trained with. In addition, the model trained as such may be used as a single generic model which can be applied to inspection SEM images obtained under different conditions and having different characteristics. Some examples of data augmentation techniques applied to reference images are discussed below.

[0077] As an example, data augmentation techniques may be applied to create a larger dataset of reference images or training images with different pixel sizes from a single reference image having a uniform pixel size. A deep learning model may be implemented to predict defects in images having a different resolution or different pixel sizes in comparison to the reference image. The deep learning model may be trained using a data augmentation technique. In some embodiments, the reference image may comprise a high-resolution human-reviewed reference SEM image having a small pixel size. It is to be appreciated that one or more reference images may be used as well. In some disclosedembodiments, training the deep learning model may include identifying a defect location or a defect image patch and adjusting the magnification of the region of the image containing the defect image patch, thereby manually creating defect image patches with different pixel sizes. As an example, if the pixel size of a reference image in the defect patch is 2 nm, a data augmentation technique may be applied to create defect image patches having a pixel size of 3 nm, 4 nm, 5 nm, 6 nm, 8 nm, 10 nm, or any appropriate pixel size, thus creating a large dataset of defect image patches with different resolution. Applying one or more data augmentation techniques may allow the artificial neural network or the deep learning model to be trained such that it focuses on the features of a defect instead of a specific defect size, thereby enabling the use of a generic deep learning model to detect and classify defects in images with lower resolution or higher pixel sizes than the training images. As a contrast, currently existing automatic defect classification (ADC) algorithm may be trained by using one or more classifiers for corresponding pixel sizes and SEM acquisition conditions.

[0078] As another example, data augmentation techniques may be applied to create a larger dataset of reference images or training images with different SNRs or “averaging.” One of several ways to obtain a high-resolution SEM image may include repetitively scanning the same region on the wafer multiple times. The images acquired from repeated scans may be stacked and averaged so that the white noise is filtered, and random blurring may be avoided. While the image “high-averaging” process may generate high SNR images, some drawbacks of this technique include long image acquisition time, increased likelihood of beam damage to the sample, and low throughput. A deep learning model may be implemented to predict defects in images having a different SNR in comparison to the SNR of the reference image. The deep learning model may be trained using a data augmentation technique. In some embodiments, the reference image may comprise a high-resolution human-reviewed reference SEM image having a high SNR or “high-average scan.” It is to be appreciated that one or more reference images may be used as well. In some disclosed embodiments, training the deep learning model may include applying a data augmentation technique by adding white noise or gaussian blurring with different kernel sizes. Adding noise, as used in the context of this disclosure and in image processing, refers to the deliberate replacement of a pixel to be different than what it may originally represent. Some non-limiting examples of noise include white noise (consistent level of noise added uniformly), salt and pepper noise (randomly changing pixels to completely white or completely black), speckle noise (uniformly adding), Gaussian blurring, Poisson blurring, etc. Adding or deliberately injecting noise into a high-resolution reference image may allow imitation of the low-resolution images. Using such a data augmentation technique may enable the artificial neural network or the deep learning model to identify the latent features of a defect instead of directly performing detection based on superficial gray level characteristics. By doing so, the trained deep learning model, during the inference phase, may be configured to detect defects in low- resolution images (e.g., low SNR images) as well as high-resolution images, rendering itself as ageneric deep learning model. As a contrast, in the existing ADC algorithms, human intervention is generally relied upon to train different classifiers for corresponding SEM imaging conditions, thereby rendering the technique inefficient, error-prone, or inconsistent.

[0079] In some embodiments, the deep learning model may be trained based on information associated with a defect in a reference image. A reference image may contain one or more types of defects, one or more occurrences of a single defect, or one or more occurrences of multiple defects. In some embodiments, a reference image may be a high-resolution, human-reviewed image containing ground truth information. Information associated with a defect may include, but is not limited to, a location of the defect, a size of the defect, the type of defect, or gray scale variations at the defect location with reference to non-defect locations in the reference image.

[0080] Inspection images (e.g., input image 502) and reference images may have a number of characteristics including, but not limited to, resolution, pixel size, sharpness, signal-to-noise ratio (SNR), brightness, contrast, or gray-level. In some embodiments, a characteristic of an image, e.g., a SEM image, may depend on the acquisition conditions, which can be adjusted based on the material being inspected, application or the desired analysis, or based on the information sought from the image. A set of acquisition conditions may refer to operating parameters of the SEM apparatus such as extracting voltage, accelerating voltage, beam current, beam size, beam focus, detector settings, deflection voltages, among other things. Adjusting one or more operating parameters of the SEM apparatus may result in a change in one or more characteristics of the image obtained. For example, the size of the probing beam may be adjusted to create a smaller probe spot on the sample, which may produce a high-resolution or a small pixel size image.

[0081] In some embodiments, a characteristic of the reference image used to train the deep learning model may be dissimilar to a corresponding characteristic of the inspection image from which a defect may be detected, identified, or classified. In such cases, the deep learning model may be trained by applying one or more data augmentation techniques to the reference images to match or substantially match with the characteristics of the inspection image. Implementation of data augmentation techniques to expand the dataset of reference images without capturing the reference images for every condition or for every characteristic corresponding to the inspection images may significantly reduce the computation and acquisition workload, data storage requirements, among other advantages. In existing defect detection and classification techniques, however, a large dataset of reference images having similar characteristics or acquisition condition as the inspection images is created and used for comparison and defect detection, rendering the technique inefficient, time-consuming, and unreliable, in addition to requiring human effort to obtain the reference images.

[0082] In some embodiments, a characteristic of the reference image used to train the deep learning model may be substantially similar to a corresponding characteristic of the inspection image from which a defect may be detected, identified, or classified. In this context, the term “substantiallysimilar” refers to a negligibly small difference in a characteristic of the reference image and the inspection image, such that the difference may not negatively impact the output of defect detection and classification by the trained deep learning model. As an example, the SNR of a pixel of a reference image obtained under an acquisition condition may be less than one standard deviation within the SNR of a corresponding pixel of an input image. It is to be appreciated that the acceptable difference between the characteristic values of a reference image and an inspection image may vary based on the characteristic being compared, or as determined by the user.

[0083] At step 520, input image 502 may be received by or fed into a deep learning network configured to implement the trained deep learning model from step 510. The trained deep learning model may be configured to detect one or more defects in input image 502. In some embodiments, the trained deep learning model may comprise an object detection algorithm to detect a defect in input image 502. The trained deep learning model, upon implementation of the object detection algorithm, may be configured to determine whether a defect is present in input image 502. In some embodiments, training the deep learning model may include labeling the data. The trained deep learning model may be a single-stage, reference-free deep learning model. As used herein, the term “reference-free model” refers to the capability of the single-stage deep learning model to identify defects and classify the defects based on the extracted features without comparison with a reference image.

[0084] The trained deep learning model may further be configured to extract one or more features of the detected defect in input image 502. In an object detection and classification framework, a deep learning model or an algorithm may be used to generate regions of interest or region proposals. The region proposals may be a large set of bounding boxes spanning the full image. In the feature extraction step, visual features may be extracted for each bounding box. The extracted features may be evaluated to determine whether and which objects are present in the region proposals based on the extracted visual features. In some embodiments, feature extraction may be performed to reduce a variable sized image to a fixed set of visual features.

[0085] The trained deep learning model may further be configured to classify the detected defect in input image 502 based on the extracted features. In some embodiments, the trained deep learning model may compare one or more extracted features of input image 502 to corresponding feature of a reference image having similar or substantially similar characteristics. The reference image used for comparison may be an as-obtained image having similar characteristics or an augmented image generated by using a data augmentation technique such that the characteristics are substantially similar to the input image (e.g., input image 502) being inspected for defects.

[0086] At step 530, the trained deep learning model may be configured to generate an output or a plurality of outputs 531, 533, 535, 537, or 539, comprising information associated with the detected defect. The deep learning model may be trained to generate an output comprising informationassociated with the detected defect. In some embodiments, an output (e.g., output 531) may comprise a portion of input image 502 containing the detected defect, a classification or an identification of the type of defect, and a confidence score associated with the detection and classification of the detected defect. A confidence score, as used herein, refers to an evaluation metric of the object detection algorithm such as the deep learning algorithm. The confidence score shows the probability of an object being detected and correctly classified as a certain defect type in an image by the algorithm, and is generally represented by a percentage value. In other words, a confidence score reflects the likelihood that the candidate object falls into a certain classifying category, and may be the confidence level of the classifier. Generally, the confidence score is higher for tighter bounding boxes having a high Intersection over Union (loU) values and a higher classification accuracy. An loU value refers to a coefficient of similarity for two sets of data. In the context of object detection, the loU is a ratio of the area of overlap of the predicted bounding box and the ground truth bounding box to the area of their union..

[0087] In some embodiments, the trained deep learning model may be configured to generate a list of detected defects and the corresponding confidence scores. It is to be appreciated that although Fig. 5 illustrates five exemplary outputs 531, 533, 535, 537, and 539, any number of outputs may be generated, as appropriate or as desired. In some embodiments, outputs 531, 533, 535, 537, and 539, may comprise a list of predicted defects (e.g., defects with a confidence score), a location of the defect, a size of the defect, or a type of the defect, or other suitable associated information.

[0088] Image analysis method (e.g., method 500) using a single-stage deep learning framework for defect detection and defect classification may have numerous advantages over existing image analysis method (e.g., process 300) in improving accuracy, efficiency, and throughput of defect inspection, among other things. Image analysis process using a single-stage trained deep learning model may have some or all of the advantages discussed herein:1. Reference-free defect detection - In the proposed image analysis method, a trained deep learning model is configured to extract one or more features of defects detected in the inspection image and classify the defects based on the extracted features, instead of comparing the inspection image to a reference image captured under similar imaging conditions and having substantially similar characteristics. Because the trained deep learning model is used to extract features directly from the inspection image, the proposed image analysis method eliminates the step of comparing it with a reference image, thereby simplifying the process and enhancing the inspection throughput and efficiency.2. Alignment-free defect detection - One of several key steps in the existing defect inspection methods (e.g., die-to-die or die-to-database inspection) include accurate alignment of the reference image with the inspection image before detection of defects can be performed, and in some cases the images may have to be realigned at later stagesas well. The accuracy in alignment relies on human intervention or extensive computation resources, rendering the process inefficient and time-consuming. The proposed image analysis method, however, completely eliminates the alignment step, thereby reducing the complexity and enhancing the efficiency and throughput of defect inspection.3. Generic and adaptive deep learning model - In addition to being reference-free and alignment-free, the proposed single-stage trained deep learning model is adaptive to changes in inspection image characteristics such as pixel size, resolution, brightness, sharpness, gray level, among other characteristics. The proposed deep learning model is trained using data augmentation techniques to create a larger dataset of reference images having different characteristics without capturing images under different imaging conditions to mimic the characteristics of the inspection image. A generic and adaptive deep learning model may be applied on images with different characteristics and acquired under different imaging conditions, rendering the model as light-weight, thereby reducing inference time compared to the existing defect inspection techniques.4. Scale robustness across different pixel sizes - During SEM condition tuning, adjusting pixel size is commonly performed to achieve high inspection throughput. The pixel size selected needs to guarantee a minimal Signal-to-Noise-Ratio (SNR) that satisfies an acceptable capture rate or true positives. The proposed deep learning model is trained using reference images collected with higher resolution and smaller pixel sizes, together with larger pixel sizes, to obtain a trained and adaptive deep learning model that can be applied to inspection SEM images with different pixel size settings.5. Noise and contrast robustness across different imaging conditions - In training the proposed deep learning model, one or more signal models may be generated which map the imaging conditions (e.g., averages, focus drift, etc.) to image characteristics such as, but not limited to, brightness, contrast, noise level, sharpness, among other things. During training data preparation, such signal models are applied to inject variations in image characteristics to improve robustness of the model against different imaging conditions.6. Nuisance robustness against pattern variations - Pattern variation is one of several contributors to high nuisance rate in defect inspection processes. Such variation cannot be suppressed by normal smoothing techniques if the pattern variation is larger than the shot noise. In the proposed deep learning model training, typical pattern variation samples are generated based on the data type. For example, for line / space patterns, nuisance samples capture variation caused by line edge roughness (LER) or non-uniformity of critical dimensions (CD). Such samples together with their nuisance label are included in the training sample pool as negative samples, which can then be applied to reject false positives and improve capture rates, thereby improving detection efficiency and accuracy.7. Higher defect inspection throughput - The conventional image analysis method using manually labeled reference images requires collection of a large set of high-resolution images of regions of interest at different conditions and scan parameters. The defect classification step also relies on human effort to select the features and tune parameters for building effective classifiers. Such process steps are inefficient, error-prone, and increase the inference time. On the other hand, the proposed image analysis based on single-stage, reference-free and alignment-free deep learning model, performs defect inspection and classification of an inspection image based on extracted features from the inspection image and may enhance the overall throughput and accuracy.

[0089] Fig. 6 illustrates a flowchart representing steps of an exemplary method 600 of image analysis with examples of input and output images, consistent with embodiments of the present disclosure. In some embodiments, one or more steps of image analysis method 600, also referred to herein as method 600, may be performed by EBI system 100 of Fig. 1, a processor associated with EBI system 100, or a defect detection and classification system (e.g., system 400 of Fig. 4). One or more steps, not illustrated in Fig. 6, may be added, deleted, edited, or ordered differently, as appropriate.

[0090] At step 610, analogous to step 510 of Fig. 5, an input image 602 may be obtained from an inspection system or an image acquisition system (e.g., image acquirer of controller 109 of EBI system 100 of Fig. 1). In some embodiments, input image 602 may be an image to be inspected or an inspection image (analogous to inspection image 302 of Fig. 3). Input image 602 may comprise a charged-particle beam image acquired using a charged-particle beam system such as, for example, a SEM system. In some embodiments, input image 502 may be a SEM image, or a backscattered electron (BSE) image, or an image formed using charged-particles originating from a region of a sample upon irradiation with a primary charged-particle beam or a probing beam of electrons.

[0091] At step 620, analogous to step 520 of Fig. 5, a single-stage trained deep learning model may be implemented to detect and classify defects in input image 602 based on features extracted from input image 602. The extracted features from input image 602 may be processed to detect one or more defects identified in a reference image during training of the deep learning model.

[0092] At step 630, analogous to step 530 of Fig. 5, the trained deep learning model may be configured to generate outputs 631 and 633 comprising information associated with the detected defect in step 620. Outputs 631 and 633 may be analogous to one of outputs 531, 533, 535, 537, or 539 of Fig. 5.

[0093] Reference is now made to Figs. 7A, 7B, and 7C, which illustrate exemplary defect signature maps 700A, 700B, and 700C, respectively, obtained using different image analysis methods, consistent with embodiments of the present disclosure. Defect signature map 700A of Fig. 7A represents a mapping of defects detected by the proposed method (e.g., method 500 of Fig. 5 ormethod 600 of Fig. 6) disclosed herein. Defect signature map 700B of Fig. 7B represents a mapping of defects detected by an exemplary existing defect detection technique before classification. Defect signature map 700C of Fig. 7C represents a mapping of defects detected by an exemplary existing defect detection technique after human-engineered or human-reviewed classification.

[0094] Defect signature map 700B of Fig. 7B illustrates a defect density mapping as detected by an existing defect detection technique, without classification of detected defects. Defect signature map 700C of Fig. 7C illustrates a human-reviewed and classified mapping of defects shown in defect signature map 700B. Defect signature map 700C may comprise ground truth information containing a negligibly small number of false positives. In other words, defect signature map 700C may be obtained by rejecting the false positives based on human engineered classification. In the context of object detection, a false positive (FP) refers to a classification of a bounding box as an incorrect detection of a non-existing object or a misplaced detection of an existing object. In some embodiments, a false positive may also be referred to herein as a nuisance detection, and the rate of occurrences of such FPs may be referred to as nuisance rate. In this context, a true positive (TP) refers to a classification of a bounding box as a correct detection of a ground truth bounding box or a correct detection of an object being present in the detected bounding box. In some embodiments, a true positive may also be referred to as a capture detection, and the rate of occurrences of such TPs may be referred to as capture rate.

[0095] As shown in Fig. 7A, defect density map 700 A, which represents the defect density mapping obtained by the disclosed single-stage deep learning model for detection and classification of defects, closely matches the ground truth mapping of defects shown in defect density map 700C. The level of similarity between defect density maps 700A and 700C is an indication of the effectiveness of the disclosed image analysis method using a single-stage deep learning model.

[0096] Reference is now made to Fig. 8, which shows a data graph 800 comparing the capture rates and nuisance rates for high resolution data used in training the deep learning model for different defect detection techniques, consistent with embodiments of the present disclosure. Data graph 800 compares a conventional die-to-die inspection technique using the high-resolution data and a deep learning-based defect inspection and classification technique relying on the high-resolution data. Data graph 800 compares the capture rate, i.e., occurrences of true positives (TPs) as a function of nuisance rates, i.e., occurrences of false positives (FPs) for the die-to-die inspection (D2D) and the deep learning-based defect inspection and classification (Deep DIC) techniques using the high-resolution data. As illustrated, data graph 800 includes nuisance rate or nuisance count along the x-axis (on a logarithmic scale) and capture rate or capture count along the y-axis (on a linear scale).

[0097] For D2D, high-resolution data may include high-resolution images for alignment and / or comparison to detect anomalies in an inspection image. For Deep DIC technique, the deep learning model may be trained using high-resolution data such as including high-resolution reference imagesof defects. As shown in data graph 800, for the proposed Deep DIC technique (denoted by the dashdash line), for capture counts < 60, the number of FPs (nuisance count) may be between 10%-30%, however, for capture counts > 60, the FPs may be less than 10% of the TPs (capture count), demonstrating the effectiveness and reliability of the defect detection and classification technique.

[0098] Referring to Fig. 9, which shows a data graph 900 comparing the capture rates and nuisance rates for low-resolution data used in training the deep learning model for different defect detection techniques, consistent with embodiments of the present disclosure. Data graph 900 compares a conventional die-to-die inspection technique relying on the high-resolution data and a deep learningbased defect inspection and classification technique using the low-resolution data. Data graph 900 compares the capture rate as a function of nuisance rates for the D2D inspection and the Deep DIC techniques using the low-resolution data. As illustrated, data graph 900 includes nuisance rate or nuisance count along the x-axis (on a logarithmic scale) and capture rate or capture count along the y- axis (on a linear scale).

[0099] As shown in data graph 900, for the proposed Deep DIC technique (denoted by the dash-dash line), the number of FPs (nuisance count) may be less than 5% of the TPs (capture count), demonstrating the effectiveness and reliability of the defect detection and classification technique using low-resolution data.

[0100] Fig. 10 is a process flowchart representing an exemplary method 1000 for analyzing an inspection image, consistent with embodiments of the present disclosure. The steps of method 1000 can be performed by a system (e.g., system 400 of FIG. 4) executing on or otherwise using the features of a computing device, e.g., controller 109 of Fig. 1. It is appreciated that the illustrated method 1000 can be altered to modify the order of steps and to include additional steps.

[0101] At step 1010, an inspection image (e.g., input image 502 of Fig. 5) is acquired using a charged-particle beam system such as, a SEM. An inspection may be acquired by an image acquisition system (e.g., image acquirer of controller 109 of EBI system 100 of Fig. 1). In some embodiments, inspection image may be a SEM image, or a backscattered electron (BSE) image, or an image formed using charged-particles originating from a region of a sample upon irradiation with a primary charged-particle beam or a probing beam. In some embodiments, inspection image may be acquired under an acquisition condition of a charged-particle beam system, for example, a SEM. The acquisition condition may comprise a set of operating parameters including, but not limited to, probe beam current, brightness, contrast, accelerating voltage, lens focus, probe -beam spot size, among other things. The acquisition condition under which inspection image may be acquired may cause the inspection image to have a characteristic such as, a gray level, a brightness, a contrast, a pixel size, a resolution, a signal-to-noise ratio (SNR), a sharpness, etc. In some embodiments, inspection image may be a low-resolution SEM image, or a high-resolution SEM image.

[0102] At step 1020, a deep learning model is trained to detect and classify defects in the inspection image based on one or more features of defects extracted from a reference image. A deep learning model may include an artificial neural network, or a single-stage deep learning framework configured to jointly detect and classify defects. Training the deep learning model may include providing one or more reference images to the deep learning model. In some embodiments, reference images may comprise high-resolution SEM images having a small average pixel size, or low-resolution SEM images having a large average pixel size, a low-brightness SEM image, a high-brightness SEM image, a low-contrast SEM image, a high-contrast SEM image, a low SNR SEM image, a high SNR SEM image, or a combination thereof. Reference image may be a SEM image of a region of a wafer, or a sample, or a substrate, or a region of interest on the wafer (e.g., a semiconductor substrate) including one or more defects. Training a deep learning model may further include providing information associated with the one or more defects in the reference image including, but not limited to, a defect location, a defect size, a defect type, among other things, to the deep learning model.

[0103] In some embodiments, training a deep learning model may further include performing one or more data augmentation techniques such as, but not limited to, noise-adding, pixel sizing, downsampling, up-sampling, image flipping, image scaling, image cropping, image rotation, adjusting contrast, adjusting brightness, adjusting saturation, performing translation, among other techniques, to reference images. Applying data augmentation techniques to reference images may allow creation of a larger dataset of reference images from a small set of captured images. In some embodiments, data augmentation techniques may be applied to reference images having a characteristic (by virtue of being captured under a condition) to create one or more reference images having a different characteristic.

[0104] At step 1030, the trained deep learning model is used to determine whether there is a defect in the inspection image received. Upon determining that there is a defect, the trained deep learning model is configured to detect a defect in the inspection image. In some embodiments, the trained deep learning model may comprise an object detection algorithm to detect a defect in the inspection image.

[0105] At step 1040, the trained deep learning model is further configured to extract one or more features of the detected defect in the inspection image. In an object detection and classification framework, a deep learning model or a deep learning algorithm is used to generate regions of interest or region proposals. The region proposals may be a large set of bounding boxes spanning the full image. In the feature extraction step, visual features are extracted for each bounding box. The extracted features are evaluated to determine whether and which objects are present in the region proposals based on the extracted visual features.

[0106] At step 1050, the trained deep learning model is further configured to classify the detected defect in the inspection image based on the extracted features. In some embodiments, the trained deep learning model compares one or more extracted features of the inspection image to correspondingfeature of a reference image having similar or substantially similar characteristics. The reference image used for comparison may be an as-obtained image having similar characteristics or an augmented image generated by using a data augmentation technique such that the characteristics are substantially similar to the inspection image being inspected for defects.

[0107] In some embodiments, the trained deep learning model is configured to generate an output list of defects and information associated with the detected or predicted defects in the list. The information associated with the detected defects may include, a defect type, a defect location, a defect size, or a confidence score of the prediction of defects.

[0108] A non-transitory computer readable medium may be provided that stores instructions for a processor of a controller (e.g., controller 109 of Fig. 1) for controlling the electron beam tool and for performing automatic defect detection and classification (e.g., as shown in Figs. 3, 5, 6, and 10), consistent with embodiments in the present disclosure. For example, instructions may include acquiring an inspection image of a sample, training a deep learning model to detect and classify defects in the inspected image. Common forms of non-transitory media include, for example, a floppy disk, a flexible disk, hard disk, solid state drive, magnetic tape, or any other magnetic data storage medium, a Compact Disc Read Only Memory (CD-ROM), any other optical data storage medium, any physical medium with patterns of holes, a Random Access Memory (RAM), a Programmable Read Only Memory (PROM), and Erasable Programmable Read Only Memory (EPROM), a FLASH- EPROM or any other flash memory, Non-Volatile Random Access Memory (NVRAM), a cache, a register, any other memory chip or cartridge, and networked versions of the same.

[0109] The embodiments may further be described using the following clauses:1. A method for defect inspection, the method comprising: acquiring a scanned inspection image of a region of a wafer, using a charged-particle beam system; and training a deep learning model based on a scanned reference image of the region of the wafer, the trained deep learning model configured to: detect a defect in the acquired scanned inspection image; after detection, extract one or more features of the detected defect; and classify the detected defect based on the one or more extracted features.2. The method of clause 1, wherein the deep learning model is trained based on information associated with a corresponding defect in the scanned reference image.3. The method of clause 2, wherein a characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.4. The method of clause 2, wherein a characteristic of the scanned reference image is substantially similar to a corresponding characteristic of the scanned inspection image.5. The method of any one of clauses 3 and 4, wherein the characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.6. The method of any one of clauses 3-5, wherein training the deep learning model comprises modifying at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.7. The method of clause 6, wherein modifying the at least one characteristic comprises performing a data augmentation technique.8. The method of any one of clauses 1-7, wherein the trained deep learning model is further configured to generate an output comprising information associated with the detected defect.9. The method of clause 8, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.10. The method of any one of clauses 1-9, wherein the trained deep learning model comprises a single-stage deep learning network.11. A defect inspection apparatus, comprising: a memory storing a set of instructions; and at least one processor configured to execute the set of instructions to cause the apparatus to perform operations comprising: acquiring a scanned inspection image of a region of a wafer using a charged-particle beam; and training a deep learning model based on a scanned reference image of the region of the wafer, the trained deep learning model configured to: detect a defect in the acquired scanned inspection image; after detection, extract one or more features of the detected defect; and classify the detected defect based on the one or more extracted features.12. The apparatus of clause 11, wherein the operations further comprise training the deep learning model based on information associated with a corresponding defect in the scanned reference image.13. The apparatus of clause 12, wherein a characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.14. The apparatus of clause 12, wherein a characteristic of the scanned reference image is substantially similar to a corresponding characteristic of the scanned inspection image15. The apparatus of any one of clauses 13 and 14, wherein the characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.16. The apparatus of any one of clauses 12-15, wherein the operations further comprise training the deep learning model by modifying at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.17. The apparatus of clause 16, wherein modifying the at least one characteristic comprises performing a data augmentation technique.18. The apparatus of any one of clauses 11-17, wherein the trained deep learning model is further configured to generate an output comprising information associated with the detected defect.19. The apparatus of clause 18, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.20. The apparatus of any one of clauses 11-19, wherein the trained deep learning model comprises a single-stage deep learning network.21. A non- transitory computer readable medium that stores a set of instructions that is executable by at least one processor of a computing device to cause the computing device to perform operations for inspecting defects in an inspection image, the operations comprising: acquiring a scanned inspection image of a region of a wafer, using a charged-particle beam system; and training a deep learning model based on a scanned reference image of the region of the wafer, the trained deep learning model configured to: detect a defect in the acquired scanned inspection image; after detection, extract one or more features of the detected defect; and classify the detected defect based on the one or more extracted features.22. The non-transitory computer readable medium of clause 21, wherein the operations further comprise training the deep learning model based on information associated with a corresponding defect in the scanned reference image.23. The non-transitory computer readable medium of clause 22, wherein a characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.24. The non-transitory computer readable medium of clause 22, wherein a characteristic of the scanned reference image is substantially similar to a corresponding characteristic of the scanned inspection image.25. The non-transitory computer readable medium of any one of clauses 23 and 24, wherein the characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.26. The non-transitory computer readable medium of any one of clauses 23-25, wherein the operations further comprise training the deep learning model by applying a data augmentation technique to adjust the characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.27. The non-transitory computer readable medium of clause 26, wherein the data augmentation technique comprises noise-adding, de-noising, pixel sizing, down-sampling, or up-sampling data associated with the characteristic of the scanned reference image to be adjusted.28. The non-transitory computer readable medium of any one of clauses 21-27, wherein the trained deep learning model is further configured to generate an output comprising information associated with the detected defect.29. The non-transitory computer readable medium of clause 28, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.30. The non-transitory computer readable medium of any one of clauses 21-29, wherein the trained deep learning model comprises a single-stage deep learning network.31. A computer-implemented method for detecting and classifying defects in an image using a deep learning model, the method comprising: providing a scanned inspection image of a region of a wafer, acquired using a charged- particle beam system; detecting, by a trained deep learning model, a defect in the scanned inspection image; extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image; classifying the detected defect based on the extracted one or more features; and generating an output comprising information associated with the detected defect, wherein the trained deep learning model is trained based on information associated with a defect in a scanned reference image of the region of the wafer, and wherein at least one characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.32. The computer-implemented method of clause 31, wherein the at least one characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.33. The computer-implemented method of any one of clauses 31 and 32, wherein training the deep learning model comprises modifying the at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.34. The computer-implemented method of any one of clause 33, wherein modifying the at least one characteristic comprises performing a data augmentation technique.35. The computer-implemented method of any one of clauses 31-34, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.36. The computer-implemented method of any one of clauses 31-35, wherein the trained deep learning model comprises a single-stage deep learning network.37. A defect inspection apparatus, comprising: a memory storing a set of instructions; and at least one processor configured to execute the set of instructions to cause the apparatus to perform operations comprising: acquiring a scanned inspection image of a region of a wafer using a charged-particle beam; detecting, by a trained deep learning model, a defect in the scanned inspection image; extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image; classifying the detected defect based on the extracted one or more features; and generating an output comprising information associated with the detected defect, wherein the trained deep learning model is trained based on information associated with a defect in a scanned reference image of the region of the wafer, and wherein at least one characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.38. The apparatus of clause 37, wherein the at least one characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.39. The apparatus of any one of clauses 37 and 38, wherein the at least one processor is configured to execute the set of instructions to cause the apparatus to further perform training the deep learning model by modifying the at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.40. The apparatus of clause 39, wherein the at least one processor is configured to execute the set of instructions to cause the apparatus to further perform a data augmentation technique to modify the at least one characteristic of the scanned reference image.41. The apparatus of any one of clauses 37-40, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score, of the detected defect.42. The apparatus of any one of clauses 37-41, wherein the trained deep learning model comprises a single-stage deep learning network.43. A non-transitory computer readable medium that stores a set of instructions that is executable by at least one processor of a computing device to cause the computing device to perform operations for inspecting defects in an inspection image, the operations comprising: acquiring a scanned inspection image of a region of a wafer using a charged-particle beam; detecting, by a trained deep learning model, a defect in the scanned inspection image; extracting, by the trained deep learning model, one or more features of the detected defect in the scanned inspection image; classifying the detected defect based on the extracted one or more features; and generating an output comprising information associated with the detected defect, wherein the trained deep learning model is trained based on information associated with a defect in a scanned reference image of the region of the wafer, and wherein at least one characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.44. The non-transitory computer readable medium of clause 43, wherein the operations further comprise training the deep learning model by applying a data augmentation technique to adjust the at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.45. The non-transitory computer readable medium of clause 44, wherein the data augmentation technique comprises noise-adding, de-noising, pixel sizing, down-sampling, or up-sampling data associated with the at least one characteristic of the scanned reference image to be adjusted.46. The non-transitory computer readable medium of any one of clauses 43-45, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.

[0110] Block diagrams in the figures may illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer hardware or software products according to various exemplary embodiments of the present disclosure. In this regard, each block in a schematic diagram may represent certain arithmetical or logical operation processing that may beimplemented using hardware such as an electronic circuit. Blocks may also represent a module, segment, or portion of code that comprises one or more executable instructions for implementing the specified logical functions. It should be understood that in some alternative implementations, functions indicated in a block may occur out of the order noted in the figures. For example, two blocks shown in succession may be executed or implemented substantially concurrently, or two blocks may sometimes be executed in reverse order, depending upon the functionality involved.Some blocks may also be omitted. It should also be understood that each block of the block diagrams, and combination of the blocks, may be implemented by special purpose hardware-based systems that perform the specified functions or acts, or by combinations of special purpose hardware and computer instructions.

[0111] It will be appreciated that the embodiments of the present disclosure are not limited to the exact construction that has been described above and illustrated in the accompanying drawings, and that various modifications and changes may be made without departing from the scope thereof. The present disclosure has been described in connection with various embodiments, other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

[0112] The descriptions above are intended to be illustrative, not limiting. Thus, it will be apparent to one skilled in the art that modifications may be made as described without departing from the scope of the claims set out below.

Claims

CLAIMS1. A defect inspection apparatus, comprising: a memory storing a set of instructions; and at least one processor configured to execute the set of instructions to cause the apparatus to perform operations comprising: acquiring a scanned inspection image of a region of a wafer using a charged- particle beam; and training a deep learning model based on a scanned reference image of the region of the wafer, the trained deep learning model configured to: detect a defect in the acquired scanned inspection image; after detection, extract one or more features of the detected defect; and classify the detected defect based on the one or more extracted features.

2. The apparatus of claim 1, wherein the operations further comprise training the deep learning model based on information associated with a corresponding defect in the scanned reference image.

3. The apparatus of claim 2, wherein a characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.

4. The apparatus of claim 2, wherein a characteristic of the scanned reference image is substantially similar to a corresponding characteristic of the scanned inspection image.

5. The apparatus of claim 3, wherein the characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.

6. The apparatus of claim 2, wherein the operations further comprise training the deep learning model by modifying at least one characteristic of the scanned reference image to substantially match the corresponding characteristic of the scanned inspection image.

7. The apparatus of claim 6, wherein modifying the at least one characteristic comprises performing a data augmentation technique.

8. The apparatus of claim 1, wherein the trained deep learning model is further configured to generate an output comprising information associated with the detected defect.

9. The apparatus of claim 8, wherein the information associated with the detected defect comprises an estimated location, an estimated size, an estimated type, or a defect confidence score of the detected defect.

10. The apparatus of claim 1, wherein the trained deep learning model comprises a single-stage deep learning network.

11. A non-transitory computer readable medium that stores a set of instructions that is executable by at least one processor of a computing device to cause the computing device to perform operations for inspecting defects in an inspection image, the operations comprising: acquiring a scanned inspection image of a region of a wafer, using a charged-particle beam system; and training a deep learning model based on a scanned reference image of the region of the wafer, the trained deep learning model configured to: detect a defect in the acquired scanned inspection image; after detection, extract one or more features of the detected defect; and classify the detected defect based on the one or more extracted features.

12. The non-transitory computer readable medium of claim 11, wherein the operations further comprise training the deep learning model based on information associated with a corresponding defect in the scanned reference image.

13. The non-transitory computer readable medium of claim 12, wherein a characteristic of the scanned reference image is dissimilar to a corresponding characteristic of the scanned inspection image.

14. The non-transitory computer readable medium of claim 12, wherein a characteristic of the scanned reference image is substantially similar to a corresponding characteristic of the scanned inspection image.

15. The non-transitory computer readable medium of claim 13, wherein the characteristic of the scanned reference image comprises a resolution, a pixel size, an acquisition condition, or a gray-level of the scanned reference image.

Images