Mid-axis detection of yield-limiting defects using 3D stacked charged particle beam inspection images

Abstract

An apparatus for detecting defects in a sample 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: acquire a plurality of images, each image being a location to be inspected; segment each of the plurality of images; stack the plurality of segmented images into a three-dimensional (3D) volume; calculate the central axis of the 3D volume; generate a central axis skeleton of the 3D volume based on the calculated central axis; and evaluate the central axis skeleton to determine whether any detected defects are present in the sample.

Description

Cross-references to related applications

[0001] This application claims priority to EP application 23216077.0, filed on December 12, 2023, which is incorporated herein by reference in its entirety. Technical Field

[0002] The embodiments provided herein relate to defect detection, and more particularly to a method for detecting defects by determining the central axis of a 3D stack of an image inspected by a charged particle beam. Background Technology

[0003] Defect inspection in semiconductor manufacturing is typically performed using image processing algorithms called "difference detectors." These algorithms are based on image comparison and involve several well-defined steps: denoising the image (using kernel filtering), aligning the image, identifying differences, and applying a threshold. In the case of repeating patterns, each image is compared to its own shifted version. In the case of non-repeating patterns, the image must be compared to images taken at the same location on different dies (or, alternatively, to a layout file in GDS format or a "golden image," i.e., an image known to be accurate).

[0004] This image processing algorithm can use rigid registration (i.e., shifting the image along the X-axis, along the Y-axis, or by rotation), which can be sensitive to local alignment errors. Such misalignment can generate shifts between the target and reference images, triggering erroneous defect indications. Using deformable registration (distorting the target image to match the distortion in the reference image) is complex because defects may include missing features in the data, leading to artifacts in the difference images and compromising the robustness of the algorithm. Summary of the Invention

[0005] Some embodiments provide an apparatus for performing operations to detect defects in a sample. The apparatus may include a memory storing an instruction set and at least one processor configured to execute the instruction set to cause the apparatus to: acquire a plurality of images, each image being a location to be inspected; segment each of the plurality of images; stack the plurality of segmented images into a three-dimensional (3D) volume; calculate the central axis of the 3D volume; generate a central axis skeleton of the 3D volume based on the calculated central axis; and evaluate the central axis skeleton to determine whether any detected defects are present in the sample.

[0006] Other advantages of the embodiments of this disclosure will become apparent from the following description taken in conjunction with the accompanying drawings, in which certain embodiments of the invention are illustrated and exemplified. Attached Figure Description

[0007] The above and other aspects of this disclosure will become more apparent from the description of exemplary embodiments in conjunction with the accompanying drawings.

[0008] Figure 1 This is a schematic diagram illustrating an example method for detecting defects using image differences.

[0009] Figure 2 This is a schematic diagram illustrating an example charged particle beam detection (CPBI) system consistent with some embodiments of the present disclosure.

[0010] Figure 3 The illustrations are consistent with some embodiments of this disclosure. Figure 2 A schematic diagram of an example charged particle beam tool, which is part of an example charged particle beam inspection system.

[0011] Figure 4 The illustration shows embodiments consistent with this disclosure. Figure 2 A schematic diagram of an example multi-beam tool, which is part of an example charged particle beam inspection system.

[0012] Figure 5 This is a schematic diagram illustrating an example neural network consistent with some embodiments of the present disclosure.

[0013] Figure 6 This is a schematic diagram illustrating an example of a polygonal skeletonization consistent with some embodiments of the present disclosure.

[0014] Figure 7 This is a schematic diagram illustrating an example process for creating a central skeletonization of a stacked CPBI image, consistent with some embodiments of this disclosure.

[0015] Figure 8 This is a schematic diagram illustrating another example process for creating a central skeletonization of a stacked CPBI image, consistent with some embodiments of this disclosure.

[0016] Figure 9 This is a flowchart illustrating an example method for detecting defects in a sample, consistent with embodiments of this disclosure. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings, wherein, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The implementations set forth in the following description of the exemplary embodiments do not represent all implementations. Rather, they are merely examples of apparatuses and methods consistent with various aspects of the disclosed embodiments as set forth in the appended claims. For example, although some embodiments are described in the context of the use of electron beams, this disclosure is not limited thereto. Other types of charged particle beams (e.g., including protons, ions, muons, or any other charged particles) can be similarly applied. Furthermore, other imaging systems, such as optical imaging, photoelectric detection, X-ray detection, ion detection, etc., can be used.

[0018] For clarity, the relative dimensions of components in the accompanying drawings may be exaggerated. Throughout the following description of the drawings, the same or similar reference numerals refer to the same or similar components or entities, and only differences relative to the various embodiments are described. As used herein, unless specifically stated otherwise, the term "or" covers all possible combinations except those that are impractical. For example, if an illustrative component may include A or B, then unless specifically stated otherwise or impractical, the component may include A or B or A and B. As a second example, if an illustrative component may include A, B, or C, then unless specifically stated otherwise or impractical, 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.

[0019] Electronic devices consist of circuits formed on a semiconductor material called a substrate. Semiconductor materials can include, for example, silicon, gallium arsenide, indium phosphide, or silicon germanium. Many circuits can be formed together on the same silicon wafer and are called integrated circuits or ICs. The size of these circuits has drastically decreased, allowing more circuits to be mounted on the substrate. For example, the IC chip in a smartphone can be as small as a fingernail but can include more than 2 billion transistors, each less than 1 / 1000th the size of a human hair.

[0020] Manufacturing these ICs, which have extremely small structures or components, is a complex, time-consuming, and expensive process, typically involving hundreds of individual steps. Even an error in one step can result in a defective IC that renders it unusable. Therefore, one goal of the manufacturing process is to avoid such defects in order to maximize the number of functional ICs manufactured in the process; that is, to increase the overall yield of the process.

[0021] A key component of increasing yield is monitoring the chip manufacturing process to ensure it produces a sufficient number of functional integrated circuits. One way to monitor this process is to inspect it at various stages of chip circuit structure formation. This inspection can be performed using scanning charged particle microscopy (SCPM). For example, SCPM could be a scanning electron microscope (SEM). SCPM can be used to image these extremely small structures, essentially taking "photographs" of the wafer structure. This image can be used to determine if the structure is correctly formed in the correct location. If a defect is found, the process can be adjusted to make the defect less likely to recur.

[0022] SCPMs (such as SEMs) work similarly to cameras. Cameras take pictures by receiving and recording the intensity of light reflected or emitted from a person or object. SCPMs take "pictures" by receiving and recording the energy or number of charged particles (such as electrons) reflected or emitted from a wafer structure. Typically, these structures are fabricated on a substrate (such as a silicon substrate) placed on a platform (called a stage) for imaging. Before taking such a "picture," a beam of charged particles can be projected onto the structure, and as the charged particles reflect or are emitted ("leaving") from the structure (e.g., from the wafer surface, from a structure below the wafer surface, or both), the SCPM's detectors can receive and record the energy or number of these charged particles to generate an examination image. To take such a "picture," the charged particle beam can scan the wafer (e.g., in a line-by-line or zigzag pattern), and the detectors can receive the leaving charged particles from the area under the projection of the charged particle beam (called a "beam spot"). The detectors can receive and record the leaving charged particles from each beam spot individually and combine the information recorded for all beam spots to generate an examination image. Some SCPMs use a single beam of charged particles (called a "single-beam SCPM," such as a single-beam SEM) to capture a single "photograph" to generate an inspection image, while others use multiple beams of charged particles (called "multi-beam SCPMs," such as multi-beam SEMs) to capture multiple "sub-photographs" of the wafer in parallel and stitch them together to generate an inspection image. By using multiple beams of charged particles, SCPMs can provide more beams of charged particles to the structure to obtain these multiple "sub-photographs," thereby allowing more charged particles to leave the structure. Therefore, the detector can simultaneously receive more leaving charged particles and generate inspection images of the wafer structure with greater efficiency and faster speed.

[0023] As the physical dimensions of IC components continue to shrink, the accuracy and yield of defect detection become increasingly important. Measurement tools can be used to determine whether an IC is manufactured correctly by identifying multiple defects on each wafer, including defects at different levels of detail, such as pattern level, image (field of view) level, die level, area of ​​interest level, or wafer level.

[0024] Figure 1 This is a schematic diagram illustrating an example method 100 for detecting defects using image differences. A target image 102 includes the defect shown in magnified portion 104. A reference image 106 includes the defect shown in magnified portion 108. A difference detector image processing algorithm is applied to the target image 102 and the reference image 106 to determine any differences between the two images. The target image 102 and the reference image 106 are aligned (e.g., via rigid registration), and the differences between the target image 102 and the reference image 106 are determined to generate a difference image 110. If any differences exist between the target image 102 and the reference image 106, these differences will be represented as detected defects 112. In some implementations, the determination of the differences between the target image 102 and the reference image 106 may be based on the number of differing pixels or the intensity difference between pixels at the same location. For example, a threshold difference in the number of pixels or a threshold difference in the intensity may be required before considering the difference as a defect.

[0025] Embodiments of this disclosure provide a method for detecting defects in a sample. According to some embodiments of this disclosure, multiple images are acquired, each image representing a location to be inspected for a potential defect (e.g., a partition location on a wafer die). The images are binarized into black and white pixels. Removing any grayscale from the images helps to make them easier to analyze. The binarized images are stacked into a three-dimensional (3D) volume. The central axis of the 3D volume is calculated, and a central axis skeleton of the 3D volume is generated based on the calculated central axis. Defects in the sample can be detected by inspecting the central axis skeleton. Defects in the skeleton are detected by discontinuities (e.g., fractures) or bifurcations (e.g., branches) in the skeleton.

[0026] Figure 2 An exemplary charged particle beam inspection (CPBI) system 200 consistent with some embodiments of this disclosure is illustrated. The CPBI system 200 can be used for imaging. For example, the CPBI system 200 can use an electron beam for imaging. Figure 2 As shown, the CPBI system 200 includes a main chamber 201, a loading / locking chamber 202, a beam tool 204, and an Equipment Front End Module (EFEM) 206. The beam tool 204 is located within the main chamber 201. The EFEM 206 includes a first loading port 206a and a second loading port 206b. The EFEM 206 may include multiple additional loading ports. The first loading port 206a and the second loading port 206b receive wafer front-opening transfer cassettes (FOUPs) containing wafers to be inspected (e.g., one or more semiconductor wafers made of (multiple) other materials) or samples (the terms "wafer" and "sample" are used interchangeably). A "batch" is a plurality of wafers that can be processed as a single load.

[0027] One or more robotic arms (not shown) in EFEM 206 can transport the wafer to loading / locking chamber 202. Loading / locking chamber 202 is connected to a loading / locking vacuum pump system (not shown), which removes gas molecules from loading / locking chamber 202 to achieve a first pressure below atmospheric pressure. After reaching the first pressure, one or more robotic arms (not shown) can transport the wafer from loading / locking chamber 202 to main chamber 201. Main chamber 201 is connected to a main chamber vacuum pump system (not shown), which removes gas molecules from main chamber 201 to achieve a second pressure below the first pressure. After reaching the second pressure, the wafer is inspected by a beam tool 204. Beam tool 204 can be a single-beam system or a multi-beam system.

[0028] The controller 209 is electronically connected to the beam tool 204. The controller 209 can be a computer capable of performing various controls of the CPBI system 200. Although in Figure 2 The controller 209 is shown outside the structure including the main chamber 201, the loading / locking chamber 202 and the EFEM 206, but it should be understood that the controller 209 may be part of the structure.

[0029] In some embodiments, controller 209 may include one or more processors (not shown). A processor may be a general-purpose or specific electronic device capable of manipulating or processing information. For example, a processor may include any number of central processing units (or “CPU”), graphics processing units (or “GPU”), optical processors, programmable logic controllers, microcontrollers, microprocessors, digital signal processors, intellectual property (IP) cores, programmable logic arrays (PLAs), programmable array logic (PALs), general-purpose array logic (GALs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), system-on-a-chip (SoCs), application-specific integrated circuits (ASICs), and any other type of circuit capable of data processing. A processor may also be a virtual processor, comprising one or more processors distributed across multiple machines or devices coupled via a network.

[0030] In some embodiments, controller 209 may also include one or more memories (not shown). Memory can be a general-purpose or specific electronic device capable of storing processor-accessible (e.g., via a bus) code and data. For example, memory can include any number of random access memory (RAM), read-only memory (ROM), optical disc, magnetic disk, hard disk, solid-state drive, flash drive, secure digital card (SD card), memory stick, compact flash (CF) card, or any combination of any type of storage device. Code can include an operating system (OS) and one or more applications (or "applications") for a specific task. Memory can also be virtual memory, which includes one or more memories distributed across multiple machines or devices coupled via a network.

[0031] Figure 3 An example imaging system 300 consistent with some embodiments of the present disclosure is illustrated. Figure 3 The beam-making tool 204 can be configured for use in the CPBI system 200. The beam-making tool 204 can be a single-beam device or a multi-beam device. For example... Figure 3 As shown, the beam tool 204 includes a motorized sample stage 301 and a wafer holder 302 supported by the motorized sample stage 301 to hold the wafer 303 to be inspected. The beam tool 204 also includes an objective lens assembly 304, a charged particle detector 306 (including charged particle sensor surfaces 306a and 306b), an objective aperture 308, a converging lens 310, a beam-limiting aperture 312, a gun aperture 314, an anode 316, and a cathode 318. In some embodiments, the objective lens assembly 304 may include a modified oscillating decelerated immersion objective (SORIL) including a pole shoe 304a, a control electrode 304b, a deflector 304c, and an excitation coil 304d. The beam tool 204 may additionally include an energy-dispersive X-ray spectrometer (EDS) detector (not shown) to characterize the material on the wafer 303.

[0032] A primary charged particle beam 320 (or simply "primary beam 320") is emitted from the cathode 318 by applying an accelerating voltage between the anode 316 and the cathode 318. The primary beam 320 passes through a gun aperture 314 and a beam-limiting aperture 312, both of which determine the size of the charged particle beam entering a converging lens 310, which resides below the beam-limiting aperture 312. The converging lens 310 focuses the primary beam 320 before it enters the objective aperture 308 to set the size of the charged particle beam before it enters the objective assembly 304. A deflector 304c deflects the primary beam 320 to facilitate beam scanning on the wafer. For example, in a scanning process, the deflector 304c can be controlled to sequentially deflect the primary beam 320 to different positions on the top surface of the wafer 303 at different time points, providing data for image reconstruction of different portions of the wafer 303. Furthermore, the deflector 304c can be controlled to deflect the primary beam 320 to different sides of the wafer 303 at specific locations at different time points, providing data for stereoscopic image reconstruction of the wafer structure at those locations. Further, in some embodiments, the anode 316 and cathode 318 can generate multiple primary beams 320, and the beam tool 204 can include multiple deflectors 304c to simultaneously project multiple primary beams 320 onto different portions / sides of the wafer, providing data for image reconstruction of different portions of the wafer 303.

[0033] Excitation coil 304d and pole piece 304a generate a magnetic field that begins at one end of pole piece 304a and terminates at the other end. A portion of wafer 303 scanned by primary beam 320 can be immersed in the magnetic field and become charged, which in turn generates an electric field. The electric field reduces the energy of the primary beam 320 impacting the surface of wafer 303 before colliding with it. Control electrode 304b, electrically isolated from pole piece 304a, controls the electric field on wafer 303 to prevent micro-bulging of wafer 303 and ensure proper beam focusing.

[0034] Upon receiving the primary beam 320, a secondary charged particle beam 322 (or "secondary beam 322") (such as a secondary electron beam) can be emitted from this portion of the wafer 303. The secondary beam 322 can form a beam spot on the sensor surfaces 306a and 306b of the charged particle detector 306. The charged particle detector 306 can generate a signal (e.g., voltage, current, etc.) representing the intensity of the beam spot and provide this signal to the image processing system 350. The intensity of the secondary beam 322 and the resulting beam spot can vary depending on the external or internal structure of the wafer 303. Moreover, as discussed above, the primary beam 320 can be projected onto different locations on the top surface of the wafer or onto different sides of the wafer at specific locations to generate secondary beams 322 (and thus beam spots) of varying intensities. Therefore, by mapping the intensity of the beam spot to the position of the wafer 303, the processing system can reconstruct an image reflecting the internal or surface structure of the wafer 303.

[0035] Imaging system 300 can be used to inspect wafer 303 on motorized sample stage 301 and includes beam tool 204, as discussed above. Imaging system 300 may also include image processing system 350, which includes image acquirer 360, storage device 370, and controller 209. Image acquirer 360 may include one or more processors. For example, image acquirer 360 may include a computer, server, mainframe, terminal, personal computer, any kind of mobile computing device, or a combination thereof. Image acquirer 360 can be connected to detector 306 of beam tool 204 via a medium such as an electrical conductor, fiber optic cable, portable storage medium, IR, Bluetooth, Internet, wireless network, radio, or a combination thereof. Image acquirer 360 can receive signals from detector 306 and can construct an image. Image acquirer 360 can thus acquire an image of wafer 303. Image acquirer 360 can also perform various post-processing functions, such as generating contours, overlaying indicators on the acquired image, etc. Image acquirer 360 can perform adjustments such as brightness and contrast of the acquired image. Storage device 370 may be a storage medium such as a hard disk, cloud storage device, random access memory (RAM), or other types of computer-readable storage. Storage device 370 may be coupled to image acquirer 360 and may be used to store scanned original image data as the original image, post-processed images, or other images that facilitate processing. Image acquirer 360 and storage device 370 may be connected to controller 209. In some embodiments, image acquirer 360, storage device 370, and controller 209 may be integrated into a single control unit.

[0036] In some embodiments, the image acquirer 360 may acquire one or more images of a sample based on imaging signals received from the detector 306. The imaging signals may correspond to a scanning operation for imaging charged particles. The acquired image may be a single image comprising multiple imaging regions. The single image may be stored in the storage device 370. The single image may be a raw image that can be divided into multiple regions. Each of the regions may include an imaging region containing features of the wafer 303.

[0037] Figure 4 The illustration shows an example multi-beam tool 204 (also referred to herein as apparatus 204) and an image processing system 490 consistent with embodiments of the present disclosure, which image processing system 490 can be configured for use in EBI system 100 ( Figure 2 Used in ).

[0038] The beam tool 204 includes a charged particle source 402, a gun aperture 404, a converging lens 406, a primary charged particle beam 410 emitted from the charged particle source 402, a source conversion unit 412, multiple beams 414, 416, and 418 of the primary charged particle beam 410, a primary projection optics system 420, a motorized wafer stage 480, a wafer holder 482, multiple secondary charged particle beams 436, 438, and 440, a secondary optics system 442, and a charged particle detection device 444. The primary projection optics system 420 may include a beam splitter 422, a deflection scanning unit 426, and an objective lens 428. The charged particle detection device 444 may include detection sub-regions 446, 448, and 450.

[0039] The charged particle source 402, the gun aperture 404, the converging lens 406, the source conversion unit 412, the beam splitter 422, the deflection scanning unit 426, and the objective lens 428 can be aligned with the primary optical axis 460 of the device 204. The secondary optical system 442 and the charged particle detection device 444 can be aligned with the secondary optical axis 452 of the device 204.

[0040] Charged particle source 402 can emit one or more charged particles, such as electrons, protons, ions, muons, or any other charged particles. In some embodiments, charged particle source 402 can be an electron source. For example, charged particle source 402 can include a cathode, extractor, or anode, wherein primary electrons can be emitted from the cathode and extracted or accelerated to form a primary charged particle beam 410 (in this case, a primary electron beam) with a cross (virtual or real) 408. For ease of explanation and without ambiguity, electrons are used as examples in some descriptions herein. However, it should be noted that any charged particle can be used in any embodiment of this disclosure, not limited to electrons. The primary charged particle beam 410 can be visualized as being emitted from the cross 408. The aperture 404 can block peripheral charged particles of the primary charged particle beam 410 to reduce the Coulomb effect. The Coulomb effect can lead to an increase in the probe spot size.

[0041] The source conversion unit 412 may include an image forming element array and a beam-limiting aperture array. The image forming element array may include an array of micro-deflectors or microlenses. The image forming element array can form multiple parallel images (virtual or real) intersecting 408 with multiple beam waves 414, 416, and 418 of the primary charged particle beam 410. The beam-limiting aperture array can limit the multiple beam waves 414, 416, and 418. Although in Figure 4 Three beams 414, 416, and 418 are shown, but embodiments of this disclosure are not limited thereto. For example, in some embodiments, device 204 may be configured to generate a first number of beams. In some embodiments, the first number of beams may be in the range of 1 to 1000. In some embodiments, the first number of beams may be in the range of 200 to 500. In some embodiments, device 204 may generate 400 beams.

[0042] The converging lens 406 can focus the primary charged particle beam 410. The currents of the beams 414, 416, and 418 downstream of the source conversion unit 412 can be changed by adjusting the focusing capability of the converging lens 406 or by changing the radial dimensions of the corresponding beam-limiting apertures within the beam-limiting aperture array. The objective lens 428 can focus the beams 414, 416, and 418 onto the wafer 430 for imaging, and can form multiple probe spots 470, 472, and 474 on the surface of the wafer 430.

[0043] Beam splitter 422 can be a Wien filter type beam splitter that generates electrostatic dipole fields and magnetic dipole fields. In some embodiments, if they are applied, the force exerted by the electrostatic dipole fields on the charged particles (e.g., electrons) of beam waves 414, 416, and 418 on the charged particles can be substantially equal in amplitude and opposite in direction to the force exerted by the magnetic dipole fields on the charged particles. Therefore, beam waves 414, 416, and 418 can pass directly through beam splitter 422 with zero deflection angle. However, the total dispersion of beam waves 414, 416, and 418 generated by beam splitter 422 can also be non-zero. Beam splitter 422 can separate secondary charged particle beams 436, 438, and 440 from beam waves 414, 416, and 418 and guide secondary charged particle beams 436, 438, and 440 toward secondary optical system 442.

[0044] The deflection scanning unit 426 can deflect beams 414, 416, and 418 to scan probe spots 470, 472, and 474 on the surface region of wafer 430. In response to the incident beams 414, 416, and 418 at probe spots 470, 472, and 474, secondary charged particle beams 436, 438, and 440 can be emitted from wafer 430. The secondary charged particle beams 436, 438, and 440 can include charged particles (electrons) with an energy distribution. For example, the secondary charged particle beams 436, 438, and 440 can be a secondary electron beam comprising secondary electrons (energy ≤ 50 eV) and backscattered electrons (energy between 50 eV and the landing energy of beams 414, 416, and 418). The secondary optical system 442 can focus secondary charged particle beams 436, 438, and 440 onto detection sub-regions 446, 448, and 450 of the charged particle detection device 444. Detection sub-regions 446, 448, and 450 can be configured to detect the corresponding secondary charged particle beams 436, 438, and 440 and generate corresponding signals (e.g., voltage, current, etc.) for reconstructing an inspection image of the structure on or below the surface region of the wafer 430.

[0045] The generated signals can represent the intensity of the secondary charged particle beams 436, 438, and 440, and can be provided to an image processing system 490 that communicates with the charged particle detection device 444, the primary projection optics system 420, and the motorized wafer stage 480. The movement speed of the motorized wafer stage 480 can be synchronized and coordinated with the beam deflection controlled by the deflection scanning unit 426, so that the movement of the scanning probe spots (e.g., scanning probe spots 470, 472, and 474) can orderly cover the region of interest on the wafer 430. The parameters of this synchronization and coordination can be adjusted to accommodate different materials of the wafer 430. For example, different materials of the wafer 430 may have different resistivity-capacitance characteristics, which may result in different signal sensitivities to the movement of the scanning probe spots.

[0046] The intensities of the secondary charged particle beams 436, 438, and 440 can vary depending on the external or internal structure of the wafer 430, thus indicating whether the wafer 430 contains defects. Furthermore, as discussed above, beams 414, 416, and 418 can be projected onto different locations on the top surface of the wafer 430 or different sides of a local structure of the wafer 430 to generate secondary charged particle beams 436, 438, and 440 with varying intensities. Therefore, by mapping the intensities of the secondary charged particle beams 436, 438, and 440 to regions of the wafer 430, the image processing system 490 can reconstruct an image reflecting the characteristics of the internal or external structure of the wafer 430.

[0047] In some embodiments, the image processing system 490 may include an image acquirer 492, a storage device 494, and a controller 496. The image acquirer 492 may include one or more processors. For example, the image acquirer 492 may include a computer, server, mainframe, terminal, personal computer, any kind of mobile computing device, or a combination thereof. The image acquirer 492 may be communicatively coupled to the charged particle detection device 444 of the beam tool 204 via a medium such as an electrical conductor, fiber optic cable, portable storage medium, IR, Bluetooth, Internet, wireless network, radio, or a combination thereof. In some embodiments, the image acquirer 492 may receive signals from the charged particle detection device 444 and may construct an image. The image acquirer 492 may thus acquire an inspection image of the wafer 430. The image acquirer 492 may also perform various post-processing functions, such as generating contours, overlaying indicators on the acquired image, etc. The image acquirer 492 may be configured to perform adjustments to the brightness and contrast of the acquired image. In some embodiments, the storage device 494 may be a storage medium such as a hard disk, flash drive, cloud storage device, random access memory (RAM), or other types of computer-readable storage. Storage device 494 can be coupled to image acquirer 492 and can be used to store scanned original image data as the original image and post-processed images. Image acquirer 492 and storage device 494 can be connected to controller 496. In some embodiments, image acquirer 492, storage device 494 and controller 496 can be integrated into a single control unit.

[0048] In some embodiments, the image acquirer 492 may acquire one or more inspection images of the wafer based on imaging signals received from the charged particle detection device 444. The imaging signals may correspond to a scanning operation for performing charged particle imaging. The acquired image may be a single image comprising multiple imaging regions. The single image may be stored in the storage device 494. The single image may be an original image that can be divided into multiple regions. Each of the regions may include an imaging region containing features of the wafer 430. The acquired images may include multiple images of a single imaging region of the wafer 430 sampled multiple times over a time series. The multiple images may be stored in the storage device 494. In some embodiments, the image processing system 490 may be configured to perform image processing steps on multiple images of the same location on the wafer 430.

[0049] In some embodiments, the image processing system 490 may include measurement circuitry (e.g., an analog-to-digital converter) to obtain the distribution of detected secondary charged particles (e.g., secondary electrons). Charged particle distribution data collected during the detection time window can be combined with corresponding scan path data of beams 414, 416, and 418 incident on the wafer surface to reconstruct an image of the inspected wafer structure. The reconstructed image can be used to reveal various features of the internal or external structure of the wafer 430, and thus can be used to reveal any defects that may exist in the wafer.

[0050] In some embodiments, the charged particles can be electrons. When electrons from the primary charged particle beam 410 are projected onto the surface of the wafer 430 (e.g., probe spots 470, 472, and 474), the electrons from the primary charged particle beam 410 can penetrate the surface of the wafer 430 to a certain depth and interact with the particles of the wafer 430. Some electrons from the primary charged particle beam 410 can elastically interact with the material of the wafer 430 (e.g., in the form of elastic scattering or collision) and can be reflected or bounced off the surface of the wafer 430. Elastic interaction conserves the total kinetic energy of the interacting subjects (e.g., electrons from the primary charged particle beam 410), where the kinetic energy of the interacting subjects is not converted into other forms of energy (e.g., heat, electromagnetic energy, etc.). Such reflected electrons generated by elastic interaction can be called backscattered electrons (BSE). Some electrons from the primary charged particle beam 410 can inelastically interact with the material of the wafer 430 (e.g., in the form of inelastic scattering or collision). Inelastic interaction does not conserve the total kinetic energy of the interacting subjects, where some or all of the kinetic energy of the interacting subjects is converted into other forms of energy. For example, through inelastic interactions, the kinetic energy of some electrons in the primary charged particle beam 410 may lead to electronic excitation and transitions between material atoms. This inelastic interaction can also generate electrons that leave the surface of the wafer 430; these electrons can be referred to as secondary electrons (SEs). The yield or emission rate of BSEs and SEs depends, for example, on the material being examined and the landing energy of the electrons from the primary charged particle beam 410 on the material surface. The energy of the electrons in the primary charged particle beam 410 can be partly determined by its accelerating voltage (e.g., ...). Figure 4 The accelerating voltage between the anode and cathode of the charged particle source 402 is imparted. The number of BSE and SE may be more or less (or even the same) than the injected electrons of the primary charged particle beam 410.

[0051] Consistent with some embodiments of this disclosure, a computer-implemented method for training a machine learning model for defect detection may include acquiring training data, which includes inspection images of fabricated integrated circuits (ICs) and IC design layout data. As used herein, the acquisition operation may refer to any operation that receives, ingests, acknowledges, obtains, acquires, retrieves, receives, reads, accesses, collects, or is used for input data. As used herein, inspection images may refer to images obtained from charged particle inspection devices (e.g., Figure 2 System 200 or Figure 3 The image is generated by the inspection process performed by the system 300. For example, the inspection image may be generated by... Figure 3The image processing system 350 generates an SCPM image. The fabrication of an IC in this disclosure can refer to an IC fabricated on a sample (e.g., a wafer) in a semiconductor manufacturing process (e.g., a photolithography process). For example, the fabricated IC can be fabricated in the die of the sample. As used herein, IC design layout data can refer to data representing the design layout of the IC. In some embodiments, the design layout data can include a design layout file in GDS format (e.g., a GDS layout file). The design layout file can be visualized (also “rendered”) as a 2D image presenting the IC layout (referred to herein as a “rendered image”). The rendered image can include various geometric features of the IC (e.g., vertices, edges, corners, polygons, holes, bridges, vias, etc.).

[0052] In some embodiments, the IC design layout data may include an image rendered based on the IC's GDS clip data (e.g., a rendered image). As used herein, the IC's GDS clip data may refer to the design layout data of the IC to be fabricated in a die, having a GDS format. In some embodiments, the IC design layout data may include only the IC's design layout file (e.g., GDS clip data). In some embodiments, the IC design layout data may include only a rendered image of the IC. In some embodiments, the IC design layout data may include only a gold image of the IC. In some embodiments, the design layout data may include any combination of the IC's design layout file, gold image, and rendered image.

[0053] Consistent with some embodiments of this disclosure, a computer-implemented method for detecting defects in the central axis using 3D stacked images may further include training a machine learning model using the acquired training data. In some embodiments, the machine learning model may be trained by a computer hardware system. In some embodiments, as described elsewhere in this disclosure, the training data may include a known, computed central axis skeleton.

[0054] In some embodiments, machine learning can be used to generate inspection images, reference images, or other images associated with device 200, device 300, or device 204. For example, in some embodiments, the machine learning system can be used with, for example... Figures 2 to 4 The controller 209, image processing system 350, image acquirer 360, storage device 370, image processing system 490, image acquirer 492, or storage device 494 operate in association. In some embodiments, machine learning can be used in defect detection methods, such as... Figure 9Method 900. In some embodiments, the machine learning system may include a discriminative model. In some embodiments, the machine learning system may include a generative model. For example, learning can have two types of mechanisms: discriminative learning, which can be used to create classification and detection algorithms, and generative learning, which can be used to actually create models, in extreme cases, which can render images. For example, as further described below, a generative model can be configured to generate images from design clips that resemble corresponding locations on the wafer in an SCPM image. This can be performed by 1) training the generative model with the design clips and associated actual SCPM images from these locations on the wafer; and 2) feeding the model design clips in inference mode at locations where a simulated SCPM image is needed. Such simulated images can be used as reference images, for example, in die-to-database inspections.

[0055] If the model(s) include one or more discriminative models, then the discriminative models(s) can have any suitable architecture or configuration known in the art. A discriminative model (also called a conditional model) is a class of models in machine learning used to model the dependency of an unobserved variable “y” on an observed variable “x”. Within a probabilistic framework, this is achieved by modeling a conditional probability distribution P(y|x), which can be used to predict y based on x. Unlike generative models, discriminative models may not allow the generation of samples from the joint distribution of x and y. However, for tasks such as classification and regression that do not require a joint distribution, discriminative models may yield superior performance. On the other hand, generative models are generally more flexible than discriminative models in expressing dependencies in complex learning tasks. Furthermore, most discriminative models are inherently supervised and cannot be easily extended to unsupervised learning. The specific details of the application ultimately determine the suitability of choosing between discriminative and generative models.

[0056] Generative models can generally be defined as models that are inherently probabilistic. In other words, a "generative" model is not a model that performs forward simulation or a rule-based approach; therefore, it may not be necessary to model the physics of the process involved in generating the actual image or output (which is generating the simulated image or output). Instead, a generative model can be learned based on a suitable training dataset (because its parameters can be learned). Such a generative model can offer many advantages for the embodiments described herein. Furthermore, generative models can be configured with deep learning architectures because they can include multiple layers that can perform various algorithms or transformations. The number of layers included in a generative model can depend on the specific use case. For practical purposes, a suitable range of layers is from two to dozens.

[0057] Deep learning is a type of machine learning. Machine learning can often be defined as a form of artificial intelligence (AI) that gives computers the ability to learn without explicit programming. Machine learning focuses on developing computer programs that can learn, grow, and change when exposed to new data. Machine learning explores the research and construction of algorithms that can learn from and predict data—making data-driven predictions or decisions by building models from sample inputs. These algorithms overcome the limitations of strictly adhering to static program instructions.

[0058] The machine learning described herein can be further performed as described in Sugiyama's "Introduction to Statistical Machine Learning" (Morgan Kaufman, 2016, p. 534); Jebara's "Discriminative, Generative, and Imitative Learning" (MIT, 2002, p. 212); and Hand et al.'s "Principles of Data Mining (Adaptive Computation and Machine Learning)" (MIT, 2001, p. 578), which are incorporated herein by reference as if fully stated herein. The embodiments described herein can also be configured as described in those references.

[0059] In some embodiments, a machine learning system may include a neural network. For example, the model may be a deep neural network with a set of weights that models the world based on data fed to it for training. Neural networks can generally be defined as a computational approach that loosely models how the biological brain solves problems by axonally connected clusters of biological neurons based on a relatively large set of neural units. Each neural unit is connected to many other neural units, and the influence of these links on the activation state of the connected neural units can be either coercive or inhibitory. These systems are self-learning and trained, rather than explicitly programmed, and excel in domains where solutions or feature detection are difficult to express in traditional computer programs.

[0060] Neural networks typically consist of multiple layers, with signal paths traversing from front to back. The goal of neural networks is to solve problems in the same way as the human brain, although some neural networks are far more abstract. Modern neural network projects often use thousands to millions of neurons and millions of connections to function. Neural networks can have any suitable architecture or configuration known in the art.

[0061] In yet another embodiment, the model may include convolutional and deconvolutional neural networks. For example, the embodiments described herein can leverage learning concepts such as convolutional and deconvolutional neural networks to solve representation transformation problems (e.g., rendering) that are typically difficult to solve. The model may have any convolutional and deconvolutional neural network configuration or architecture known in the art.

[0062] As used in this paper, a neural network can refer to a computational model that analyzes latent relationships in an input dataset by mimicking the human brain. Similar to biological neural networks, a neural network can comprise a collection of connection units or nodes (called "neurons") constructed in different layers, where each connection (also called an "edge") can receive and send signals between neurons in neighboring layers in a manner similar to synapses in a biological brain. Signals can be any type of data (e.g., real numbers). Each neuron can receive one or more signals as input and output another signal by applying a nonlinear function to the input signals. Neurons and edges are typically weighted by corresponding weights to represent the knowledge the neural network has acquired. During the training process (similar to the learning process in a biological brain), the weights can be adjusted (e.g., by increasing or decreasing their values) to change the signal strength between neurons, thereby improving the performance accuracy of the neural network. Neurons can apply a thresholding function (called an "activation function") to the output value of their nonlinear function such that a signal is output only when the aggregated value (e.g., a weighted sum) of the output values ​​of the nonlinear function exceeds a threshold determined by the thresholding function. Neurons in different layers can transform their input signals in different ways (e.g., by applying different nonlinear functions or activation functions). The output of the final layer (called the "output layer") can be the analysis result of the neural network, such as, for example, the classification of the input dataset (e.g., in the case of image recognition), numerical results, or any type of output data used to obtain the analysis result from the input data.

[0063] As used in this article, neural network training may refer to the process of improving the accuracy of a neural network's output. Generally, training can be categorized into three types: supervised training, unsupervised training, and reinforcement training. In supervised training, a target output dataset (also known as a "label" or "ground truth") is generated based on an input dataset using methods other than neural networks. This input dataset can then be fed into the neural network to generate an output dataset that is typically different from the target output data. Based on the difference between the output data and the target output data, the weights of the neural network can be adjusted according to rules. If this adjustment is successful, the neural network can generate another output dataset that is more similar to the target output data in the next iteration using the same input data. If this adjustment is unsuccessful, the weights of the neural network can be adjusted again. After a sufficient number of iterations, the training process can be terminated according to one or more predetermined criteria (e.g., the difference between the final output data and the target output data is below a predetermined threshold, or the number of iterations reaches a predetermined threshold). The trained neural network can then be applied to analyze other input data.

[0064] In unsupervised training, neural networks are trained without any external metrics (such as labels) to identify patterns in input data, rather than generating labels for them. Typically, neural networks analyze shared properties (such as similarity and difference) and relationships between elements of input data according to one or more predefined rules or algorithms (such as principal component analysis, clustering, anomaly detection, or latent variable identification). The trained neural network can then extrapolate these identified relationships to other input data.

[0065] In reinforcement learning, neural networks are trained in a trial-and-error manner without any external metric (such as labels) to maximize the benefit of decisions. The input dataset for the neural network may differ during reinforcement training. For example, during training, reward or penalty values ​​can be determined for the output of the neural network according to one or more rules, and the weights of the neural network can be adjusted to maximize the reward (or minimize the penalty). The trained neural network can then apply its learned decision knowledge to other input data.

[0066] During the training of a neural network, a loss function (or "cost function") can be used to evaluate the output data. As used in this paper, a loss function maps the output data of a machine learning model (such as a neural network) to a real number (called "loss" or "cost") that intuitively represents the loss or error associated with the output data (e.g., the difference between the output data and the target output data). The training of a neural network can seek to maximize or minimize the loss function (e.g., by pushing the loss towards a local maximum or minimum in the loss curve). For example, one or more parameters of the neural network can be adjusted or updated to maximize or minimize the loss function. After adjusting or updating one or more parameters, the neural network can receive new input data in the next iteration of its training. The training of the neural network can be terminated when the loss function is maximized or minimized.

[0067] Through examples, Figure 5 This is a schematic diagram illustrating an example neural network 500 consistent with some embodiments of the present disclosure. For example... Figure 5 The depicted neural network 500 may include an input layer 520 that receives inputs, including inputs 510-1, ..., inputs 510-m (where m is an integer). For example, the inputs to the neural network 500 may include any structured or unstructured data (e.g., images). In some embodiments, the neural network 500 may receive multiple inputs simultaneously. For example, in... Figure 5 In this embodiment, the neural network 500 can simultaneously receive m inputs. In some embodiments, the input layer 520 can continuously receive m inputs, such that the input layer 520 receives input 510-1 in a first cycle (e.g., in a first inference) and pushes data from input 510-1 to a hidden layer (e.g., hidden layer 530-1), then receives a second input in a second cycle (e.g., in a second inference) and pushes data from the second input to a hidden layer, and so on. The input layer 520 can receive any number of inputs simultaneously, continuously, or in any way that groups the inputs.

[0068] Input layer 520 may include one or more nodes, including node 520-1, node 520-2, ..., node 520-a (where a is an integer). Nodes (also called “machine perceptrons” or “neurons”) can model the operation of biological neurons. Each node can apply an activation function to the received input (e.g., one or more of inputs 510-1, ..., input 510-m). Activation functions may include Hevise step functions, Gaussian functions, multiple quadratic functions, inverse multiple quadratic functions, sigmoid functions, rectified linear unit (ReLU) functions (e.g., ReLU6 or leaky ReLU), hyperbolic tangent (“tanh”) functions, or any nonlinear function. The output of the activation function may be weighted by weights associated with the node. Weights may include positive values ​​between 0 and 1 or any value that can scale the outputs of some nodes in the layer to more or less than the outputs of other nodes in the same layer.

[0069] like Figure 5 Further described, the neural network 500 includes multiple hidden layers, including hidden layer 530-1, ..., hidden layer 530-n (n is an integer). When the neural network 500 includes more than one hidden layer, it can be called a "deep neural network" (DNN). Each hidden layer can include one or more nodes. For example, in... Figure 5 In the hidden layer 530-1, nodes 530-1-1, 530-1-2, 530-1-3, ..., 530-1-b (where b is an integer) are included, and hidden layer 530-n includes nodes 530-n-1, 530-n-2, 530-n-3, ..., 530-nc (where c is an integer). Similar to the nodes in the input layer 520, the nodes in the hidden layers can apply the same or different activation functions to the outputs of the connected nodes from the previous layer, and the outputs from the activation functions are weighted by the weights associated with the nodes.

[0070] like Figure 5 Further described, the neural network 500 may include an output layer 540, which ultimately determines the output, including output 550-1, output 550-2, ..., output 550-d (d is an integer). The output layer 540 may include one or more nodes, including node 540-1, node 540-2, ..., node 520-d. Similar to the nodes in the input layer 520 and the hidden layers, the nodes in the output layer 540 can apply activation functions to the outputs from the connected nodes of the previous layer and weight the outputs from the activation functions through weights associated with the nodes.

[0071] Despite Figure 5The diagram depicts the nodes of each hidden layer of the neural network 500 connected to each node in its preceding and following layers (referred to as "fully connected"), but the layers of the neural network 500 can use any connection scheme. For example, one or more layers of the neural network 500 (e.g., input layer 520, hidden layers 530-1, ..., hidden layers 530-n, or output layer 540) can use convolutional schemes, sparse connection schemes, or a combination of convolutional and sparse connections between layers. Figure 5 The fully connected scheme described is connected to any connected scheme with fewer connections.

[0072] Moreover, despite Figure 5 In this diagram, the inputs and outputs of each layer of neural network 500 are depicted as propagating in the forward direction (e.g., feeding from input layer 520 to output layer 540, referred to as a "feedforward network"), but neural network 500 may additionally or alternatively use backpropagation (e.g., feeding data from output layer 540 towards input layer 520) for other purposes. For example, backpropagation can be implemented using long short-term memory nodes (LSTM). Therefore, although neural network 500 is depicted as similar to a convolutional neural network (CNN), neural network 500 may include a recurrent neural network (RNN) or any other neural network.

[0073] In some embodiments, defects can be detected by stacking multiple (segmented) CPBI images into a 3D volume and then applying a 3D midline transformation (e.g., skeletonization). The resulting skeleton can then be examined for breaks or bifurcations to identify defects. The 3D volume skeleton allows for relatively easy detection of defects by finding breaks (e.g., discontinuities) or bifurcations (e.g., branches) within the skeleton.

[0074] In summary, the method typically involves acquiring a CPBI image corresponding to the same location on a mask, segmenting the acquired image, stacking the segmented images into a 3D volume, calculating the central axis of the 3D volume, generating a central axis skeleton based on the calculated central axis, and examining the central axis skeleton to determine if there are any defects in the CPBI image, which would manifest as breaks or bifurcations in the central axis skeleton.

[0075] Figure 6 This is a schematic diagram illustrating an example of a central axis skeletonization of a polygon consistent with some embodiments of this disclosure. The central axis is defined as the set of points within the polygon closest to more than one edge. It can also (equivalently) be considered as a point that is entirely within the polygon and contacts the center of a sphere in at least two places. For example, in triangle 602, the central axis 604 is composed of angle bisectors, such as... Figure 6 As shown. A more complex polygon 610 results in a more complex central axis 612, as... Figure 6As shown. Through these examples, the central axis of the polygon forms a tree-like skeleton, composed of lines and points in 2D. In 3D volume, the skeleton becomes a collection of points, lines, and surfaces.

[0076] Figure 7 This is a schematic diagram illustrating an example process 700 for creating a central skeletonization of stacked CPBI images, consistent with some embodiments of this disclosure. A stack 702 creates multiple CPBI images 702a to 702n. Each CPBI image 702a to 702n is captured from the same region (e.g., a partition) of different dies on the wafer. This allows process 700 to examine repeating patterns in stack 702 to detect defects. Process 700 operates in a similar manner, regardless of the size of the individual images 702a to 702n in the stack. For example, each image 702a to 702n may be on a smaller scale (e.g., a partition of a die) or on a larger scale (e.g., an entire die). The way the images are created (e.g., sampled) affects how defects can be detected. For example, a defect should be fully visible within a single image, rather than spanning image boundaries (e.g., an indeterminate location because the defect is located at the edge of the image, making it appear not to be a defect based on the image).

[0077] Before creating stack 702, each of the images 702a through 702n is segmented. As used in this disclosure, "segmenting" an image may include binarizing the image to remove any grayscale and representing features in the image as black or white pixels. For example, if the sample being examined is made of two different materials, one material may be represented by black pixels, while the other material may be represented by white pixels. Other segmentation methods may be applied, such as simple thresholding or deep learning-based segmentation (which may include per-pixel classification).

[0078] Images 702a to 702n can be stacked together, for example, in the order of die exposure. Each image 702a to 702n can also be labeled (e.g., with an identifier) ​​so that if a defect is found, it can be determined which image contains the defect. Other orders of images 702a to 702n in the stack are also possible.

[0079] Then, a 3D volume 704 is created from stack 702. For illustrative purposes, "slices" of the 3D volume 704 are shown as image slices 704a to 704n, each image slice representing a top view of a slice within the 3D volume 704. Dashed lines on image slices 704a to 704n indicate the shooting position of the skeletonized 706 of the 3D volume 704. The skeletonized 706 is shown as a side view of the 3D volume 704.

[0080] like Figure 7As shown, the analyzed features include contact holes, which are shown as circles in image slices 704a to 704n. In image slice 704a, the dashed lines indicate two contact holes, which are shown as circles 708a and 708b in skeletonization 706. Because the two contact holes are separate (i.e., there is no defect connecting the two contact holes), this is reflected in skeletonization 706 as two separate line segments 710a and 710b.

[0081] In image slice 704b, the dashed lines indicate that the two contact holes are shown as circles 712 in skeletonization 706 because the two contact holes are connected by a defect. In skeletonization 706, the merging of the two contact holes is indicated by point 714, where line segments 710a and 710b indicate the merging (i.e., the defect caused by the connecting two contact holes).

[0082] In image slice 704n, the dashed line shows that one of the contact holes (the bottom center of image slice 704n) is deformed, causing the skeletonization 706 to branch from point 716 into two line segments 718a and 718b.

[0083] Skeletonization is an image or volumetric transformation technique. It's simpler to compute defects using skeletonization than to look at individual images. If skeletonization is performed on a single image, the skeleton will be a set of points or spots. When images are stacked together and if two points are connected, skeletonization includes the line between the two points, which is a clear change in the skeleton. Skeletonization is an unstable transformation, meaning that small changes in the input can lead to large changes in the output, and it can be advantageously used in defect detection. For example, detecting changes becomes trivial in the skeleton domain, but may not be easy in the image domain.

[0084] By examining the skeleton to detect breaks or bifurcations, these changes indicate defects (such as missing connections). When defects are present, the continuity of the skeleton is disrupted. The continuity of the skeleton is relatively easy to detect, for example, using connected component algorithms. For example, in Figure 7 In the image, skeletonization 706 branches at point 716 because the edge (as shown in image slice 704n) is not perfectly circular. If the edge were perfectly circular, then there would be a line in skeletonization 706 reaching that point. Because there are two separate edges (i.e., vertices), the line branches towards each vertex (as shown in line segments 718a and 718b).

[0085] If two points are connected, however slight, they will be shown in the skeletonization, which helps detect defects because defects in the original image may not be obvious. For example, a contact hole may be deformed, but if the contact hole is not connected in some way (no matter how small the connection, even a one-pixel gap), then it will not appear in the skeletonization, but will be shown as two separate points. If the same structure appears in all images of the stack, then it is assumed that the structure was correctly manufactured. The purpose of skeletonization is to find images that are different from other images in the stack (e.g., one-off manufacturing defects). For example, if an image in the middle of the stack shows a defect, it may be difficult to spot an image showing a defect under existing methods of examining all images. By using skeletonization to find differences between individual images, creating a skeletonization of the entire 3D image stack helps to identify defects.

[0086] Figure 8 This is a schematic diagram illustrating an example process 800 for creating a stacked CPBI image, consistent with some embodiments of this disclosure. (To be combined with...) Figure 7 The stack 702 described is similar to the stack 802 that creates multiple CPBI images 802a to 802n. Each CPBI image 802a to 802n is taken from the same region (e.g., a partition) of different dies on the wafer, is segmented, and the images are combined with... Figure 7 The CPBI images 702a to 702n described are stacked in a similar manner.

[0087] Then, with combination Figure 7 The volume 704 described is created from stack 802 in a similar manner to the 3D volume 804. For illustrative purposes, "slices" of the 3D volume 804 are shown as image slices 804a to 804n, each image slice representing a top view of a slice within the 3D volume 804. Dashed lines on image slices 804a to 804n indicate the shooting position of the skeletonized 806 of the 3D volume 804. The skeletonized 806 is shown as a side view of the 3D volume 804.

[0088] like Figure 8 As shown, the analyzed features include contact holes, indicated as circles in image slices 804a to 804n. In image slice 804a, dashed lines through line segment 808 indicate the presence of contact holes in skeletonization 806. In image slice 804b, contact holes are missing at the same location (indicated by dashed lines). This manifests in skeletonization 806 as discontinuities and gaps 810 (e.g., breaks in skeletonization 806) within skeletonization 806. In image slice 804n, dashed lines through line segment 812 indicate the presence of contact holes in skeletonization 806.

[0089] Figure 9This is a flowchart illustrating an example method 900 for detecting defects in a sample, consistent with embodiments of this disclosure. In some embodiments, method 900 may be... Figure 2 The image processing system 250 is executed.

[0090] In step 902, an image of the same location to be examined is acquired. For example, images 702a to 702n or 802a to 802n can be acquired from the CPBI tool.

[0091] In step 904, the acquired image is segmented. For example, the image can be segmented by binarizing it to remove any grayscale and representing the features in the image as black and white pixels.

[0092] In step 906, the segmented images are stacked into a 3D volume, such as 3D volume 704 or 804.

[0093] In step 908, the central axis of the 3D volume is calculated. The central axis is the set of points within the 3D volume that are closest to more than one edge. It can also be (equivalently) considered as a point that can be entirely within the 3D volume and touches the center of the 3D volume in at least two places.

[0094] In step 910, the central axis skeleton is calculated based on the calculated central axis, for example, skeletonization 706 or 806.

[0095] In step 912, defects are detected based on central skeletonization. For example, defects can be detected based on discontinuities (e.g., breaks) or bifurcations (e.g., branches) in the skeleton. Changes in the skeleton (such as discontinuities or bifurcations) indicate changes in the image stack of the 3D volume.

[0096] In some embodiments, machine learning (ML) may be used in conjunction with process 700, process 800, or method 900 to accelerate the entire process. In some embodiments, supervised machine learning methods suitable for surrogate modeling, such as neural networks, Gaussian processes, or support vector regression, may be used. For example, the step of computing the central skeleton can be replaced by ML prediction. In this example, the ML model is trained with a known computational skeleton. For example, the skeleton is computed using an "expensive" algorithm, and the ML model is trained to predict the skeleton from a stack of 3D images. It may be necessary to train the ML model for each type of pattern that needs to be detected, and therefore different training may be required for different features to be examined, such as contact holes, line spacing, or other features.

[0097] Non-transitory computer-readable media may be provided, the storage of which is used for controller (e.g.) Figure 2The processor instructions of the controller 209 are used to perform image inspection, image acquisition, stage positioning, beam focusing, electric field adjustment, beam bending, converging lens adjustment, activation of charged particle sources, beam deflection, and operations 700 and 800 and method 900, etc. Common forms of non-transitory media include, for example, floppy disks, flexible disks, hard disks, solid-state drives, magnetic tape or any other magnetic data storage media, compact disc read-only memory (CD-ROM), any other optical data storage media, any physical media with a hole pattern, random access memory (RAM), programmable read-only memory (PROM) and erasable programmable read-only memory (EPROM), flash memory EPROM or any other flash memory, non-volatile random access memory (NVRAM), cache, registers, any other memory chips or cassette tapes and their networked versions.

[0098] The embodiments may be further described using the following terms: 1. A non-transitory computer-readable medium storing an instruction set executable by at least one processor of a computing device to cause the computing device to perform operations for detecting defects in a sample, the operations including: Acquire multiple images, each corresponding to a location on the sample; Segment each of multiple images; Stack multiple segmented images into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Generate a 3D volumetric central axis skeleton based on computational central axis; and The central skeleton is evaluated to determine if any detected defects exist in the sample. 2. A non-transitory computer-readable medium pursuant to Clause 1, wherein each of a plurality of images is the same partition of a die on an inspected wafer. 3. A non-transitory computer-readable medium pursuant to Clause 2, wherein each of a plurality of images is a charged particle beam inspection image. 4. A non-transitory computer-readable medium according to any one of clauses 1 to 3, wherein segmentation includes binarizing each image into black pixels and white pixels. 5. A non-transitory computer-readable medium according to any one of clauses 1 to 4, wherein the stacking is based on the ordering of multiple images. 6. A non-transitory computer-readable medium pursuant to Clause 5, wherein the order of a plurality of images is based on the order in which the images were acquired. 7. A nontransitory computer-readable medium pursuant to any of Clauses 1 to 6, wherein the defect manifests in the skeleton as a discontinuity or bifurcation in the skeleton. 8. A non-transitory computer-readable medium pursuant to Clause 7, wherein evaluating the central skeleton includes using a connected component algorithm to detect discontinuities or bifurcations in the skeleton. 9. An apparatus for detecting defects in a sample, comprising: Memory, storing instruction sets; and At least one processor is configured to execute a set of instructions to cause the device to perform operations, including: Acquire multiple images, where each image represents the location to be examined; Segment each of multiple images; Stack multiple segmented images into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Generate a 3D volumetric central axis skeleton based on computational central axis; and The central skeleton is evaluated to determine if any detected defects exist in the sample. 10. The apparatus according to Clause 9, wherein each of the plurality of images is the same partition of a die on the wafer being inspected. 11. The apparatus according to Clause 10, wherein each of the plurality of images is a charged particle beam inspection image. 12. An apparatus according to any one of clauses 9 to 11, wherein segmentation includes binarizing each image into black pixels and white pixels. 13. An apparatus according to any one of clauses 9 to 12, wherein operation further includes stacking multiple segmented images based on an ordering of multiple images. 14. The apparatus according to Clause 13, wherein the ordering of a plurality of images is based on the order in which the images are acquired. 15. A device according to any one of clauses 9 to 14, wherein the defect in the skeleton is manifested as a discontinuity or bifurcation in the skeleton. 16. The apparatus according to Clause 15, wherein evaluating the central skeleton includes using a connected component algorithm to detect discontinuities or bifurcations in the skeleton. 17. A method for detecting defects in a sample, comprising: Acquire multiple images, where each image represents the location to be examined; Segment each of multiple images; Stack multiple segmented images into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Generate a 3D volumetric central axis skeleton based on computational central axis; and The central skeleton is evaluated to determine if any detected defects exist in the sample. 18. The method according to Clause 17, wherein each of the plurality of images is the same partition of the die on the wafer being inspected. 19. The method according to Clause 18, wherein each of the plurality of images is a charged particle beam inspection image. 20. The method according to any one of clauses 17 to 19, wherein segmentation includes binarizing each image into black pixels and white pixels. 21. The method according to any one of clauses 17 to 20, wherein stacking is based on the ordering of multiple images. 22. The apparatus according to Clause 21, wherein the ordering of a plurality of images is based on the order in which the images are acquired. 23. The method according to any one of clauses 17 to 22, wherein the defect in the skeleton is manifested as a discontinuity or bifurcation in the skeleton. 24. The method according to Clause 23, wherein evaluating the central skeleton includes using a connected component algorithm to detect discontinuities or bifurcations in the skeleton. 25. A method for detecting defects in a sample, comprising: Train a machine learning model on the central skeleton of the pattern to be examined; Acquire multiple images, where each image represents the same location to be examined; Segment each of multiple images; Stack multiple segmented images into a three-dimensional (3D) volume; Generate a 3D volumetric central skeleton using a trained machine learning model; and The central skeleton is evaluated to determine if any detected defects exist in the sample. 26. In accordance with the location of Clause 25, where segmentation includes binarizing each image into black pixels and white pixels. 27. The method according to Clause 25 or 26, wherein the defect in the skeleton is manifested as a discontinuity or bifurcation in the skeleton.

[0099] The block diagrams in the accompanying drawings 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 some embodiments, a non-transitory computer-readable medium is provided and may include execution of combinations. Figures 7 to 9Instructions for any one or more functions described in the diagram. In this regard, each box in the diagram may represent certain arithmetic or logical operations that can be implemented using hardware such as electronic circuits. Boxes may also represent modules, segments, or code sections, including one or more executable instructions for implementing the specified logical function. It should be understood that in some alternative implementations, the functions indicated in the boxes may not occur in the order shown in the figures. For example, two boxes shown consecutively may be executed or implemented substantially concurrently, or the two boxes may sometimes be executed in reverse order, depending on the functionality involved. Some boxes may also be omitted. It should also be understood that each box in the block diagram, and combinations of boxes, may be implemented by a system based on dedicated hardware (performing the specified function or action) or by a combination of dedicated hardware and computer instructions.

[0100] It should be understood that the embodiments of this disclosure are not limited to the precise constructions described above and illustrated in the accompanying drawings, and various modifications and changes can be made without departing from its scope. This disclosure has been described in conjunction with various embodiments, and other embodiments will be apparent to those skilled in the art in light of the specification and practice of the techniques disclosed herein. The specification and examples are intended to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

Claims

1. A non-transitory computer-readable medium storing an instruction set executable by at least one processor of a computing device to cause the computing device to perform operations for detecting defects in a sample, the operations comprising: Acquire multiple images, each corresponding to a location on the sample; Segment each of the plurality of images; The multiple segmented images are stacked into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Based on the calculated central axis, the central axis skeleton of the 3D volume is generated; as well as The central skeleton is evaluated to determine whether any detected defects exist in the sample.

2. The non-transitory computer-readable medium of claim 1, wherein each of the plurality of images is the same partition of a die on the inspected wafer.

3. The non-transitory computer-readable medium of claim 2, wherein each of the plurality of images is a charged particle beam inspection image.

4. The non-transitory computer-readable medium according to claim 1, wherein the segmentation comprises: Each image is binarized into black pixels and white pixels.

5. The non-transitory computer-readable medium of claim 1, wherein the stacking is based on the order of the plurality of images.

6. The non-transitory computer-readable medium of claim 5, wherein the ordering of the plurality of images is based on the order in which the images were acquired.

7. The non-transitory computer-readable medium of claim 1, wherein the defect in the skeleton is manifested as a discontinuity or a bifurcation in the skeleton.

8. The non-transitory computer-readable medium of claim 7, wherein evaluating the central axis skeleton comprises: The connected component algorithm is used to detect the discontinuities or bifurcations in the skeleton.

9. An apparatus for detecting defects in a sample, comprising: Memory stores instruction sets; as well as At least one processor is configured to execute the instruction set to cause the device to perform operations, the operations including: Acquire multiple images, where each image represents the location to be examined; Segment each of the plurality of images; The multiple segmented images are stacked into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Based on the calculated central axis, the central axis skeleton of the 3D volume is generated; and The central skeleton is evaluated to determine whether any detected defects exist in the sample.

10. The apparatus of claim 9, wherein each of the plurality of images is the same partition of a die on the wafer being inspected.

11. The apparatus of claim 10, wherein each of the plurality of images is a charged particle beam inspection image.

12. The apparatus of claim 9, wherein the segmentation comprises: Each image is binarized into black pixels and white pixels.

13. The apparatus of claim 9, wherein the operation further comprises: The multiple segmented images are stacked based on the sorting of the multiple images, wherein the sorting is based on the order in which the images were acquired.

14. The apparatus according to claim 9, wherein: Defects in the skeleton manifest as discontinuities or bifurcations within the skeleton; and Evaluating the central skeleton includes: using a connected component algorithm to detect the discontinuities or bifurcations in the skeleton.

15. A method for detecting defects in a sample, comprising: Acquire multiple images, where each image represents the location to be examined; Segment each of the plurality of images; The multiple segmented images are stacked into a three-dimensional (3D) volume; Calculate the central axis of the 3D volume; Based on the calculated central axis, the central axis skeleton of the 3D volume is generated; as well as The central skeleton is evaluated to determine whether any detected defects exist in the sample.

Images