Interface detection in reciprocal space

Abstract

Interface detection in reciprocal space. Methods and apparatuses determine the location of a boundary of a given material in a sample based on classification of reciprocal space images at respective locations. A diffraction image at a given location is classified to identify the material at that location. Imaging and classification at multiple locations can quickly and reliably find the boundary to sub-micron accuracy. Binary search provides acceleration. Classification is performed by a neural network. The technology is applicable to distinguishing single crystal silicon, poly silicon, and amorphous silicon, high-Z materials (tungsten) or low-Z materials (carbon), etc. The technology is integrated into an automated workflow that precisely positions further imaging, probing, or milling operations based on the located boundary. Examples and variations are disclosed.

Description

Background Technology

[0001] In the semiconductor industry, interface detection is a valuable component of sample analysis. Objects of interest are typically located at or near the boundary between two materials. Examples include blanket films or semiconductor structures built on top of bulk silicon. Conventional techniques include manual examination of survey images or machine vision techniques used for, for example, edge detection. The former is naturally unsuitable for workflow automation, while the latter can be ad-hoc and, in some cases, error-prone. First, an imaging modality is needed where the desired interface is visible as an edge or other discontinuity, which may require iterative work as new materials or device types are introduced. Specifically, analytical procedures may need to detect multiple interfaces between different material pairs. Finding a universal imaging modality can be challenging for some samples. Machine vision techniques can then locate edges well, but may fail to identify which edge is the desired one. Heuristics for selecting the correct edge can vary from sample to sample, requiring significant configuration and testing (often trial and error). Therefore, a robust, universal, and automated technique for locating material boundaries in semiconductors and other samples remains needed. Summary of the Invention

[0002] In short, the disclosed techniques use analysis of reciprocal space images to classify materials in corresponding imaging regions. By scanning a series of imaging regions (e.g., stepwise or via binary search), the location of transitions (boundaries) from one material to another can be determined quickly and reliably with an accuracy of 100 nm or higher.

[0003] Reciprocal space images can be obtained directly, for example, using a diffractometer, or indirectly, for example, by capturing physical space images and applying transformations. Each reciprocal space image can be classified, for example, using a neural network to identify the material present in the imaging region. The boundary locations of a given material can be determined across multiple imaging regions based on the classification.

[0004] The disclosed technology can be implemented as a device comprising: an imager configured to acquire multiple images of a corresponding region of a sample; and a controller. The controller can be configured to: process a reciprocal space representation of each image; and, in response to the processing, classify materials within the corresponding region among a predetermined plurality of materials. The controller can be configured to determine the coordinate position of a boundary of a first material among the plurality of materials.

[0005] The above and other objects, features and advantages of the present invention will become more apparent from the following detailed description taken with reference to the accompanying drawings. Attached Figure Description

[0006] Figure 1It is a diagram of an example device through which the disclosed technology can be implemented.

[0007] Figure 2 It is a collection of images illustrating the applications of the disclosed technology.

[0008] Figure 3 This is a diagram illustrating the use of binary search in an example of the disclosed technique.

[0009] Figure 4 It is a flowchart depicting a first example method based on the disclosed technology.

[0010] Figures 5A to 5B It is a flowchart depicting an example extension of the method based on the disclosed technology.

[0011] Figure 6 This is a diagram illustrating an example neural network that can be used to implement the disclosed techniques.

[0012] Figure 7 This is an illustration of an example of using multiple scans to locate material boundaries according to the disclosed technique.

[0013] Figure 8 Generalized examples of suitable computing environments in which the described implementation schemes, technologies, and technologies involving imaging or milling are illustrated. Detailed Implementation

[0014] Introduction

[0015] Automated analysis of semiconductors and other samples relies on the registration of analytical tools with sample features, typically material boundaries or predetermined offsets from such boundaries. Conventional techniques can be ad-hoc methods for finding material boundaries. An imaging modality suitable for one type of boundary may not be suitable for another. As the sample and composition vary, extensive trial and error may sometimes be required to find one or more suitable imaging modalities that capture all material boundaries of interest. For example, machine vision techniques used for edge detection can also be easily fooled on some samples, rejecting desired edges and favoring artifacts. When the sample changes, a machine vision algorithm that works effectively for one sample type may have a high failure rate.

[0016] The disclosed techniques rely on reciprocal space images (such as diffraction patterns) that directly provide material characterization without depending on a skilled operator or fragile machine vision heuristics. In some examples, tools such as converging beam electron diffraction (CBED) can operate with spot sizes smaller than 10 nm, allowing for very precise mapping of materials in a sample. Scanning multiple spots along a path easily identifies the location where transitions from one material to another occur. The disclosed techniques can detect changes in the CBED image that indicate changes in the underlying imaged material, such as a change from a CBED specific to single-crystal silicon to a non-specific CBED. Furthermore, a binary search along the path can find transitions within logarithmic time, meaning the time spent is logarithmically linearly related to the desired tolerance. Even further, multiple scans can be used to identify boundary orientations, trace boundaries with complex shapes, or identify multiple boundaries between corresponding material pairs within a single session.

[0017] the term

[0018] Unless otherwise expressly stated or contradicted by the context, the usage and meaning of all terms referenced in this section apply to the entire disclosure. The following terms are extended to their relevant word forms.

[0019] "Analysis" refers to operations used to characterize a sample and can include material removal, other sample preparation, imaging, probe measurements, non-contact measurements, or secondary evaluation of data obtained through any of these techniques. Analytical operations can include imaging (e.g., FIB, optical, or electron microscopy), etching (e.g., ion milling), delay, electron backscattering analysis, electron microscopy, mass spectrometry, materials analysis, metrology, nanoprobes, spectroscopy, or surface preparation. The equipment or instrument used to perform this operation is called an "analytical instrument," "analyzer," or "tool." Non-limiting examples of tools that can be used in conjunction with the disclosed techniques include milling machines, etchers, electron microscopes, electron spectrometers, optical microscopes, or optical spectrometers. Specifically, some tools of interest herein include plasma focused ion beam (PFIB), scanning electron microscopy (SEM), or fluorescence microscopy.

[0020] The term "beam" refers to the directed flow of particles or energy. Common beams of interest in this disclosure are particle beams, such as electron beams or ion beams (including plasma-focused ion beams); or optical beams, such as laser beams used for fluorescence excitation. A beam can have a finite range transverse to its main longitudinal flow direction. The line connecting the centroids of two or more cross-sections of the beam is the "axis" of the beam.

[0021] "Binary search" is a one-dimensional search technique in which discontinuities can be accurately found in logarithmic time by successively bisecting intervals of discontinuity. To illustrate the boundary between carbon and silicon, if images at coordinate positions 0 μm and 64 μm show materials carbon ("C") and silicon ("Si") respectively, the C-Si boundary can be initially specified as 32 ± 32 μm. If the next image taken at 32 μm is determined to be C, the boundary can indeed be improved to 48 ± 16 μm. Successive images can show 48 μm = Si, 40 μm = Si, 36 μm = C, 38 μm = Si, so that five measurements can improve the boundary determination from 32 ± 32 μm to 37 ± 1 μm. The number of measurements (5) is equal to the logarithm (base 2) of the accuracy improvement (32).

[0022] A "boundary" is the interface between two regions that differ in the composition, structure, or orientation of their respective constituent materials. As an example, a boundary can separate a silicon region from a carbon or tungsten region. As another example, a boundary can separate a monocrystalline silicon region from an amorphous silicon region. As a further example, a boundary can separate two polycrystalline silicon grains with different orientations. As used herein, defining a boundary can be understood as a shorthand for determining the coordinate positions of one or more points on the boundary. A boundary can be an interface between two different materials (such as monocrystalline Si and amorphous Si) or an interface within a single material, such as between two polycrystalline Si grains.

[0023] "Classification" and "classification" refer to the action of assigning an item to a selection from a finite set of predetermined choices. In some examples of interest in this paper, the classified item could be an imaging region of a sample, and each of these selections could be a corresponding material. In various examples, classification can be performed by trained machine learning software, by other automated software, or by a user (e.g., using interactive software). The software program that performs the classification is called a "classifier".

[0024] A “controller” is an electronic device coupled to one or more actuators to achieve changes in physical parameters, or coupled to one or more sensors to monitor physical parameters. Some controllers may contain a microprocessor that can be programmed to execute machine-readable instructions. The description of computing devices herein generally applies to such controllers, which may also include additional electronic circuitry systems (such as filters and amplifiers). Other controllers may contain analog circuitry systems, such as filters and amplifiers, without any microprocessor.

[0025] A converged beam electron diffractometer (“CBED”) is an analyzer configured to perform electron diffraction analysis of a sample using a converged electron beam. The converged beam provides a narrower beam spot (typically ranging from 1 nm to 100 nm or 0.3 nm to 300 nm in diameter) than a parallel beam (typically at least 500 nm), improving spatial resolution. The converged beam can form a cone with a half-width of approximately 0.1–1°, which can manifest as a broadening of features observed in reciprocal space; for example, a conventional electron diffraction spot appears as a disk-like broadening. In some examples, the CBED can project the electron diffraction pattern onto a fluorescent screen and image it using a CCD camera. Other detectors and readouts can also be used.

[0026] "Coordinate position" numerically specifies the location of an associated entity in physical space, where each of one or more coordinates indicates the distance of that location in the corresponding dimension. Coordinates can be one-dimensional coordinates along a straight line or curve, two-dimensional coordinates along a plane or surface, or three-dimensional coordinates. Coordinates can include a mixture of distance and angular coordinates (e.g., in polar, cylindrical, or spherical coordinate systems), or only distance coordinates (e.g., in Cartesian coordinate systems). The coordinate position of a point on a boundary is called a "boundary position."

[0027] An "electron beam" is a directed flow of electrons.

[0028] An electron microscope is an analyzer in which a sample is illuminated by an electron beam, and the resulting particles or electromagnetic radiation are used to form a spatially resolved direct image. Scanning electron microscopy (SEM) images the surface of a sample based on reflected, secondary, or backscattered particles or radiation from one or more surfaces of the sample. Because the beam interactions detected by SEM occur at or near the surface, SEM can operate on samples of arbitrary thickness. In contrast, transmission electron microscopy (TEM) images the sample volume based on transmitted electrons (including scattered electrons). TEM operates on samples with thicknesses from about 10 nm to 150 nm, which can be mounted on grids for mechanical support and thermal conduction. TEM can provide magnifications up to and exceeding 50 million (with resolutions below 5 pm), while SEM magnifications are typically limited to about 2 million (with resolutions as low as about 0.5 nm). In this disclosure, scanning transmission electron microscopy (STEM), which uses a scanning electron beam to perform imaging of transmitted electrons, is considered both SEM and TEM. The electron beam in an electron microscope is generated in an electron gun and accelerated, focused, or redirected toward the sample chamber through a series of stages.

[0029] A “feature” is a structural or other variation that is identifiable in an image of a sample. In a direct image, a feature may include structural or material boundaries, such as components of a memory cell, transistors, grain boundaries, or interfaces between a bulk substrate layer and another layer. In reciprocal space, exemplary features may include moiré patterns, Kikuchi lines, a central spot, or a halo. Reciprocal space features can provide clues about the composition, structure, or orientation of the material in an imaging region of the sample, and variations in these features from one region to the next can be used to determine material boundaries. A “moiré pattern” is a repetitive oscillation of image intensity, sometimes associated with interference received, for example, from two sources or scatterers. A “Kikuchi line” is a bright band in an electron diffraction pattern associated with a lattice plane of the imaged sample. The position and orientation of the Kikuchi line depend on the tilt and azimuth orientation of the associated lattice plane relative to the incident electron beam. A reciprocal image may also have a “central spot” as a bounded bright region and a “halo” as a ring-shaped region with an observable intensity (per pixel) that is at least one to two times lower than the peak intensity of the central spot. The central spot is typically circular or nearly circular (e.g., the ratio of the large diameter to the small diameter is in the range of 1 to 2) and is usually located at the center of the reciprocal image (e.g., aligned with the axis of the incident beam), but these are not required, and in other examples, the central spot may be eccentric or elongated. Features may exhibit “symmetry” in the direct or reciprocal image, meaning that a single feature or a set of features remains unchanged in either (1) a finite rotation of less than 360° around an axis (called the “axis of symmetry”), or (2) a reflection in a plane (called the “plane of symmetry”). The axis of symmetry and the plane of symmetry can be considered to extend orthogonally from the image plane.

[0030] The term "imaging" refers to the process of obtaining a two-dimensional representation (referred to as an "image") of parameter values ​​on a region of interest in a sample. In some examples, the imaging parameters may be backscattered or secondary emitted from an incident particle beam (e.g., by SEM or low-current focused ion beams), but this is not required, and visible light or other parameters may also be imaged. In the disclosed examples, the image may indicate an object on or within the sample (e.g., a gate, layer, metallization, or via) or a feature of such an object (e.g., an edge).

[0031] An "imager" (sometimes, an "imaging tool") is a device capable of imaging. Some imagers use the response to an incident light beam (e.g., beam current, secondary emission, backscattering, or fluorescence) as a measurement parameter for forming an image. An imager that acquires or provides a direct image of a target sample is called a "microscope." Microscopes that use an incident electron beam or ion beam may be called an "electron microscope" (e.g., SEM, STEM, or TEM) or an "ion microscope" (e.g., FIB), respectively. An imager in which the measurement parameters include the intensity, wavelength, or another property of light may be called an "optical microscope." Some optical microscopes use an incident optical beam (e.g., for fluorescence), but this is not necessary, and other beams or diffuse or ambient illumination may also be used. Other imagers acquire or provide a reciprocal space image of a sample and may include diffractometers such as CBEDs or X-ray diffractometers. The term "charged particle imager" encompasses electron microscopes, ion microscopes, and CBEDs. Various detectors can be used to measure particles or radiation emitted from the imaged sample, including secondary electron detectors, secondary ion detectors, photodetectors, or spectrometers. The technique used to acquire the image is called an "imaging mode." Therefore, scanning electron microscopy, transmission electron microscopy, and convergent beam electron diffraction are different modes utilized by SEM, STEM, TEM, and CBED tools, respectively. In some examples, the same beam column can be used for multiple imaging modes. In further examples, the analyzer may include multiple beam columns, such as corresponding columns for electron beams, ion beams, or laser beams.

[0032] "Material" refers to the characterization of a substance present in a sample region. Characterization may include composition; for example, silicon, silicon dioxide, and carbon are three different materials. Characterization may also include structure; for example, carbon can have a hexagonal or cubic structure in graphite and diamond, respectively. Silicon can exist in single-crystal, polycrystalline, or amorphous forms. "Polycrystalline" silicon ("polycrystalline Si") can be organized into grains (typically ranging from 2 nm to 500 nm) with a regular crystalline structure within each grain and orientation variations (or another discontinuity) between adjacent grains. In this disclosure, if the composition and structure are the same, the substance in two regions is considered the same material even if the orientations are different. Therefore, adjacent polycrystalline silicon grains are considered the same material. Furthermore, a material may have internal boundaries, such as places where discontinuities exist in its structure (e.g., at grain boundaries). "Amorphous" silicon lacks a regular structure: the orientation of Si-Si bonds 1 nm or more may be unrelated. “Single-crystal” silicon (“single-crystal Si”) has a uniform crystal structure in a plane ranging from more than 1 μm to 30 cm, but may be adjacent to layers containing amorphous or polycrystalline silicon. Some materials of interest in this paper may have a composition characterized by atomic number. Silicon has an atomic number of 14. Elements with atomic numbers greater than or equal to 15 (such as tungsten) can be considered “high-Z” materials, while elements with atomic numbers less than or equal to 13 (such as carbon) can be considered “low-Z” materials. In further examples, the structure of a material can encompass the physical layout of a device or component on or within a sample. For example, a memory array may have a specific regular structure that differs from the organization of adjacent logic or processor blocks. Thus, the memory array and the logic or processor blocks can be considered as different materials that can be easily distinguished in reciprocal space (e.g., in a transformation of a direct image), with the memory array having sharp spectral peaks according to its regular layout and the logic or processor blocks having diffuse spectral density.

[0033] A “neural network” is an artificial network of “units” (or “cells”) that have connections that mimic the behavior of biological neurons and can be implemented using software programs on a computer. A common neural network can be organized into layers, where an input array of data is operated on to generate an output array at each layer. Some neural networks described herein are “convolutional neural networks” (CNNs), where at least one layer involves multiple segments of the input array and has one or more common cores to generate a corresponding output array. In examples, the disclosed techniques can be implemented on residual CNNs (ResNets). A neural network has a structure, such as layers, neuron types, and operations or paths from the input to the output array at each layer; and also has an array of coefficients or weight data to be applied to each neuron or operation. Examples of the disclosed techniques apply neural networks to, for example, the classification of materials in an imaging sample region.

[0034] A "parameter" is a quantity that can have a specific "value". While parameters typically have numerical values, this is not required, and some parameter values ​​can be logical values, strings, or data structures. Some parameters can be control parameters, for example, set by a controller to affect a physical device or physical operation. Other parameters can be sensed parameters, for example, determined by the physical environment and measured or evaluated by a controller or measuring device (e.g., to form an image or to measure or evaluate from an image).

[0035] "Physical space" refers to the two-dimensional or three-dimensional region in which a sample tangibly exists. A sample in physical space has length, width, and thickness (or alternative extent). An image in which the extent or spacing of features is proportional to their extent or spacing in physical space is called a "direct image." For illustration, a visible light camera or microscope can form a direct image of a sample.

[0036] "Reciprocal space" refers to a two- or three-dimensional range where coordinate values ​​are proportional to spatial frequencies in physical space. For illustration, consider a spatial periodicity feature A whose wavelength is twice that of another spatial periodicity feature B. In reciprocal space, the coordinate position of feature A could be half that of feature B. An image in reciprocal space is called a "reciprocal image." Diffraction tools such as CBED can generate reciprocal images directly from a sample. Reciprocal images can also be obtained by transforming a direct image (e.g., via FFT). For example, transformed images can be sensitive to the texture of the imaging surface or can allow for the identification or differentiation of materials based on spectral content.

[0037] A “region” is a continuous two-dimensional portion of the sample surface or a continuous three-dimensional portion of the sample volume.

[0038] A “region of interest” (ROI) is a portion of a sample that includes features or structures that are objects of subsequent analysis. An ROI can be defined relative to the surface of the sample (e.g., an area of ​​an imaging surface containing a specific feature, or an area where a pattern can be milled to form a sheet containing that feature) or relative to a volume (e.g., a volume containing a structure to be further analyzed). The term ROI is not intended to involve any human interest. A region can be a region of interest; however, some regions may lack specific features or structures and may not be regions of interest.

[0039] A “sample” is a physical object on which imaging, milling, or other analytical or manufacturing processes are performed. Common samples may include biological samples, multilayer electrical or electronic structures, or other material structures. Electrical or electronic samples may include semiconductor devices such as logic gates, memory cells, optoelectronic devices, passive components (including interconnects), transistors, and may be at various stages of manufacturing or disassembly. Material samples may include microelectromechanical systems (MEMS) or nanostructures. Biological samples may include cells, proteins, DNA, RNA, viruses, or fragments thereof. The disclosed techniques can be applied during sample preparation, sample characterization, or manufacturing. Samples may include a “substrate” on which cells, structures, or devices are deposited or fabricated prior to sheet preparation or other analysis.

[0040] A "scan" is the traversal of an imager across a sample, such as taking a series of images at corresponding locations. The imager can scan imaging positions along a predetermined path. In some examples, successive imaging positions can be "sequential," meaning the coordinate values ​​of the imaging positions change monotonically during the scan. For example, for each successive image, the coordinates of the imaging positions can increase in steps of 200 nm. In other examples, a binary search can be advantageously used, allowing the coordinate positions of boundaries to be rapidly reduced to predetermined tolerances. Although scanning is typically a one-dimensional scan performed along a straight or curved path, this is not necessary, and two-dimensional scans can also be used, for example, to map complex boundary profiles. Furthermore, any combination of discrete sets of one-dimensional scans, offset from each other in different directions, can be used.

[0041] A “semiconductor device” is a device comprising one or more semiconductor regions (e.g., doped or undoped silicon, germanium, III-V semiconductors, or similar materials). A “manufactured semiconductor device” can be a finished product (e.g., a product that performs its intended electronic function when a suitable electrical signal is applied), a portion thereof, or a product at an intermediate stage of manufacturing (e.g., a product on which at least one patterning operation has been performed). Semiconductor devices typically include other materials to provide insulating or conductive paths between semiconductor regions. Non-limiting examples of semiconductor devices include transistors, memory cells, and some optoelectronic devices. Memory interface circuitry may include both conductive wiring and logic gates.

[0042] "Software" refers to a computer-executable program, instructions, or associated data structure. Software can be active or inactive. In an active state, software can be loaded into memory or executed by one or more processors. In an inactive state, software can be stored on a computer-readable medium, awaiting transmission or execution. "Instructions" (or "program instructions") encode operations to be performed by the processor (sometimes along with one or more operands, operand addresses, or destination addresses). A set of instructions can be organized to perform a function. "Interactive software" refers to a software program that presents information to a user and receives commands from the user in response to the presented information. "Automation software" refers to a software program that receives information from equipment or other software programs and selects one or more actions to be performed by the equipment or other software programs without user input.

[0043] "Tolerance" is a predetermined limit on the acceptable deviation between the reported value of a parameter and its true value. In some examples, material boundaries can be defined to a tolerance of 100 nm, meaning that the reported boundary coordinates are within 100 nm of the precise location of the boundary. Tolerances can range from 1 nm to 10 mm.

[0044] The exposed master surface of the horizontal sample has an upward-pointing outward normal relative to it, for example, a common configuration from which the sample can be approached by the processing tool from above. The terms "top," "bottom," "upward," "downward," "above," "below," "horizontal," "vertical," etc., are used for convenience. The axis of the processing tool (e.g., the sight axis of an imaging tool or the beam axis of a FIB milling machine) may be substantially vertical downward onto the surface, or may be tilted relative to the normal at an angle (sometimes in the range of 40° to 60°). Those skilled in the art will understand from this disclosure that the choice of actual orientation may be altered without departing from the scope of the disclosed technology.

[0045] "Training" refers to the process of determining the values ​​(coefficients) to be applied at the neurons of a neural network, enabling the network to operate and perform its desired function. Training can be performed using a training dataset that includes images whose desired classification is known. Comparing the actual output of the trained neural network to the desired classification provides a loss function, which is adjusted by backpropagating the network, applying gradient descent, or another established technique to adjust the coefficients of each layer. As training progresses, the neural network output converges to the desired classification, and the magnitude of the loss function can decrease. When the loss function decreases below a predetermined threshold, the neural network can be considered "trained" and can be deployed to classify new images whose correct classification may not be known.

[0046] A "transformation" refers to the act of converting a direct image (or direct spatial data) into a reciprocal image (or reciprocal spatial data) or vice versa. This transformation can be performed via software (e.g., using Fast Fourier Transform (FFT) or Wavelet Transform) or physically via the imager. For example, a CBED can transform a physical structure into a reciprocal spatial image.

[0047] "User" is a person. An exemplary user may control the analysis process or equipment, or may receive the presentation or notification of data from such analysis equipment. The user may interact with the equipment, sample, or software via communication networks, computer input or output interfaces, or client software.

[0048] Example devices

[0049] Figure 1 Figure 100 illustrates an example device through which the disclosed technology can be implemented. This device delivers an imaging beam onto a sample to obtain an image, which can be classified by a controller. The imaging beam can be scanned relative to the sample surface to determine the boundary between a first material and another material.

[0050] Imager 140 may include a beam source 142, a detector 144, and a reader 146. Imaging beam source 142 may be configured to guide an incident beam 112 along axis 105 onto sample 120. Beam 112 may be scanned over sample 120 to obtain a series of images at corresponding locations on sample 120. Sample 120 may be fixed on stage 130. Detector 144 may be configured to detect scattering particles from sample 120. Signals representing the particles may be measured by reader 146.

[0051] In this example, the imager can be a SEM, STEM, TEM, or CBED, but this is not required, and another microscope or diffractometer can also be used. As shown, beam 112 is convergent, as might be used in a CBED, but this is not required, and a parallel beam can be used. A divergent beam can also be used, for example, to increase the irradiated area of ​​the sample. As shown, detector 144 can be a segmented ring detector, as can be found in a diffractometer, but this is not required, and in other examples, one or more discrete sensors, such as an Everhart Thornley detector, a microchannel plate, or a Faraday cup, can be used. Readout 146 may include one or more amplifiers and / or ammeters. Scanning of beam 112 over sample 120 can be performed by manipulating beam 112, translating stage 130, or a combination thereof.

[0052] The controller 150 may be directly or indirectly coupled to the imager beam source 142, stage 130, reader 146, or detector 144. Thus, the controller 150 can control and acquire images from the imager 140 and cause the beam 112 to scan relative to the sample 120. The controller 150 can also process the images received from the imager 140 to, for example, classify materials within the imaging regions of the sample 120 in a predetermined set of materials. In some examples, the imager 140 may directly provide the controller 150 with a reciprocal image, while in other examples, the imager 140 may provide the controller 150 with a direct image, which may be transformed into a reciprocal space representation by the controller 150 prior to classification. Based on the material classification at a series of imaging regions of the sample 120, the controller 150 can determine the coordinate positions of the boundaries of a first material in the sample 120.

[0053] Various extensions or variations of the illustrated device can be implemented within the scope of the disclosed technology. In the examples, imager 140 may be a direct space imager (such as SEM, STEM, TEM, or FIB) or a reciprocal space imager (such as CBED). The illustrated device may also support a second imaging modality, such as another electron microscopy modality, FIB, or fluorescence imaging. Controller 150 may be configured to acquire another image of sample 120 using the second imaging modality. This image may be acquired at a second coordinate position having a predetermined spatial relationship with respect to the coordinate position of the boundary. In some examples, beam 112 may be used for both imaging modalities, while in other examples, both imaging modalities may be provided in a dual-beam tool. Detector 146 may be a pixelated electron diffraction detector with at least 100 pixels (and up to one billion pixels). The current received at each pixel of detector 146 may accumulate as a charge at each pixel, and the pixel charge may be read out using pixel array readout electronics 146.

[0054] In some examples, controller 150 may be configured to identify boundaries of monocrystalline silicon, such as interfaces between a monocrystalline silicon bulk layer and one or more other materials, such as amorphous silicon, polycrystalline silicon, tungsten or another high-Z material, or carbon or another low-Z material. In some examples, controller 150 may be configured to perform binary classification, such as (1) a given first material or (2) not a given first material, while in other examples, controller 150 may be configured to classify an image as indicating any one of three or more materials.

[0055] The controller 150 can use a convolutional neural network to perform classification. The controller 150 can be configured to scan the beam 112 on the sample 120 at sufficiently close intervals to establish boundary locations within a predetermined tolerance. The tolerance can be about 100 nm, or in the range of 10 nm to 50 nm, 50 nm to 200 nm, or 200 nm to 1 μm.

[0056] The controller 150 may have a wired or wireless network interface for connecting to a remote computer or control station 152.

[0057] Example Principle

[0058] Figure 2 The images 201-205 are a set of images illustrating the operating principle of the disclosed technique. Image 201 is a reproduction of a high-angle annular dark-field (HAADF) STEM image of the sample, which is a direct image formed by high-angle forward scattered electrons. CBED images 202-205 were taken at corresponding regions 212-215 of the sample and have characteristic properties of the material in the sample region.

[0059] Starting with image 202, a bright central spot 221 can be seen, which is speckled but otherwise has a generally uniform intensity. While no obvious halo is present, some regular Kikuchi lines 223 can be seen radiating outwards from the core 221. These features could be related to the axis of the incident beam (similar to...). Figure 1 The axis 105 exhibits the characteristics of a symmetrical, regular crystal structure, where strong diffraction from the principal plane enhances the visibility of the Kikuchi lines outside the core, and there are no irregular variations to wash away the spots. Such features can be associated with, for example, bulk-layered single-crystal silicon.

[0060] Moving to image 203, a bright central spot 231 can be seen, in which Kikuchi line features 233 (bright) and 235 (dark) are both within the central spot 231 and extend significantly beyond it. The image exhibits significant asymmetry. Such features can be characteristics of a regular crystal structure lacking strong symmetry about the incident beam axis, which may be caused by the tilting of the beam or lattice. Image 203 can also be associated with single-crystal silicon.

[0061] Image 204 shows a bright central spot 241 with only moderate intensity variation and no significant intensity outside the central spot 241. Such features indicate low scattering and can be associated with low-Z amorphous materials such as carbon.

[0062] Image 205 shows a central spot 251 with low intensity and some speckling, surrounded by a distinct symmetrical diffuse halo 254. The low intensity and speckling indicate a mixture of diffractions from grains with different orientations (e.g., polycrystalline materials). (CBED spot size can typically be smaller than the grain size: signal mixing can occur between grains at the sample surface and grains with different orientations buried underneath.) The halo may be a result of strong scattering from high-Z materials. Image 205 could be associated with tungsten.

[0063] Images 202-205 illustrate how variations in material in regions 212-215 can result in significantly different features in the reciprocal space representation (such as diffraction patterns). These images are merely illustrative: other features can be used to identify a wide range of materials in addition to or as alternatives to the described features. In some examples, CBED may be advantageous due to its small spot size. Materials can be accurately identified from very small areas of the sample surface, and material boundaries can be determined at very high resolution in the ranges of 10 nm to 20 nm, 20 nm to 50 nm, 50 nm to 100 nm, or 100 nm to 200 nm. Additionally, artificial neural networks can be trained to perform this material identification with high accuracy, providing better than 90%, 95%, 98%, or 99% correct material identification in varying applications. Automated, precise boundary identification automates workflows, providing registration of the sample in the field of view of, for example, a STEM analyzer. Using this boundary identification, high-resolution TEM imaging, probing, or milling can be performed at precise locations without manual assistance.

[0064] Example applications and binary search

[0065] Figure 3 Figure 300 illustrates the use of binary search in an example of the disclosed technique. Images are acquired and processed at a series of regions to narrow down the coordinate positions of the boundaries within predetermined tolerances.

[0066] 1. The disclosed technology offers advantages.

[0067] Image 310 is a reproduction of a high-angle annular dark-field (HAADF) STEM image of a sample with fins. Some bright lines in Image 310 may be STEM artifacts indicating edges, while others may indicate high-Z buffer layers. Image 310 shows a relatively simple structure of a semiconductor device; however, the device has many such bright lines varying in thickness, orientation, and length. In an automated workflow, it may be desirable to find the coordinate location of edge 312. Comparative machine vision techniques may be fooled by other edges in Image 310, resulting in low reliability (e.g., below a predetermined threshold, which may be in the range of 50% to 89%), and therefore unsuitable for automated workflows. In contrast, the disclosed technique is not fooled by STEM artifacts, thus allowing reliable determination of edge 312. That is, the disclosed technique can determine the coordinate location of edge 312 within a predetermined tolerance and with a probability greater than a predetermined threshold. The predetermined tolerance may be in the range of 1 nm to 10 μm, 10 nm to 1 μm, 30 nm to 300 nm, or about 100 nm. The predetermined threshold can be in the range of 90% to 99.999%, for example, approximately 90%, 95%, 98%, 99%, or 99.9%.

[0068] Although Figure 310 shows a relatively simple structure of a semiconductor device, the disclosed techniques can be applied to fairly complex samples, thus providing reliable automatic determination of boundary locations.

[0069] 2. Applications in automated workflows

[0070] Image 310 depicts a row of gate devices 314 formed above edge 312. In an example application, a high-resolution TEM image or another analytical procedure can be performed at a predetermined offset from edge 312 based on the coordinate position of edge 312 determined using the disclosed techniques. That is, once edge 312 is found, the layer above edge 312 can be reliably analyzed and characterized. For example, if the z-coordinate of edge 312 has been found along path 330, subsequent analysis can be reliably performed at a desired offset along path 330 (e.g., at region 351).

[0071] In another application, the orientation of edge 312 can be determined based on two scans by determining the two coordinate positions of edge 312, thereby allowing reliable localization of subsequent analysis areas (e.g., region 352) away from path 330.

[0072] In further applications, multiple boundary locations can be determined along the corresponding path and between the corresponding materials, for example, to determine the profile of gate 314. Subsequently, one or more regions can be precisely selected relative to the profile of gate 314 for subsequent analysis, such as at region 353.

[0073] 3. Accelerated Binary Search

[0074] Return to edge 312, Figure 3 The right side illustrates the coordinate position of edge 312 along path 330. Reciprocal space images are obtained at a series of positions along path 330. Representative images 321 and 325 are illustrated. Each image 321-325 is located in the z-direction according to its corresponding imaging position. For illustration, leader 364 points to region 374 where image 324 is acquired and is labeled with the corresponding z-coordinate "z4". Similar leaders indicate the z-position and imaging region (in...) of images 321-323 and 325. Figure 3 (Unmarked in the text).

[0075] Initially, as disclosed herein, images 321 and 322 at two relatively distant positions z1 and z2 can be acquired and classified. As shown in the figure, images 321 and 322 are CBED images and can be classified directly. Different features of images 321 and 322 lead to their different classifications, thereby enclosing the position of the sought edge 312 in the interval (z1, z2). Then, an image 323 can be acquired at position z3=(z1+z2) / 2 and classified. Its features may result in a classification that matches the classification of image 321, thereby indicating that edge 312 can be found within the narrowed interval (z3, z2). A fourth image 324 can be acquired at z4=(z3+z2) / 2. The classification of image 324 is different from that of image 321, although it may or may not be the same as the classification of image 322. Therefore, the position of edge 312 can be further narrowed down to (z3, z4).

[0076] So far, the interval enclosing edge 312 remains greater than a predetermined threshold, for example, T=100 nm. Therefore, the bisection continues, and a reciprocal image 325 is finally acquired, and this reciprocal image is also classified differently from image 321. Now, the interval (z3, z5) including the coordinate position of edge 312 has a span Δ=|z5-z3|<T. Since the span Δ is less than the threshold T, the binary search method can stop. The coordinate z=(z3+z5) / 2 can be returned as the position of edge 312, within the tolerance Δ / 2.

[0077] Example Method

[0078] Figure 4 is a flowchart 400 depicting a first example method in accordance with the disclosed technology. In this method, reciprocal space images at a series of sample positions are acquired and classified to identify the boundary position of a first material.

[0079] At block 410, a series of reciprocal space images can be acquired for corresponding regions of a sample. For illustration, the images may be electron diffraction images from successive positions on a sample surface. At block 420, each image can be classified to identify the material present in the corresponding region from a predetermined set of materials. Then, at block 430, the boundary position of the first material can be determined based on the classification.

[0080] Many variations and modifications can be implemented within the scope of the disclosed technology. For example, in Figure 4In the diagram, boxes 410 and 420 are illustrated as overlapping to indicate that classification can begin before image acquisition is complete. Specifically, images can be classified during acquisition, and the classification can be used to guide the selection of subsequent imaging regions. Classification at box 420 can be performed independently for each image in the image. Classification can be based on the presence or absence of a moiré pattern, the presence or absence of Kikuchi lines, the intensity of the central spot, the absolute or relative intensity of the halo, or the symmetry of one or more features of the instantaneous reciprocal image. Classification can be based on spectral properties detected in a transformation of the direct image. The sample can be a fabricated semiconductor device. The first material can be single-crystal silicon, typically used as the bulk substrate of the semiconductor device.

[0081] In the example, a series of images can be acquired using a binary search along a predetermined path at box 410. Initially, with two imaging positions on opposite sides of the sought boundary, a third image can be acquired at a location positioned along the predetermined path and between the two imaging positions. This can be repeated for additional images, acquiring the next image at a location between the most recent imaging position and another previous imaging position. Since the two previously imaging positions can be located on opposite sides of the sought boundary in any iteration, the next image can narrow the interval containing the desired boundary, and the binary search can be iteratively performed to converge to the desired boundary position within a predetermined tolerance. In some examples, the boundary position can be confirmed by taking two additional images along the predetermined path on opposite sides of the determined boundary position and classifying these additional images to confirm that exactly one of the two images indicates the presence of the first material.

[0082] exist Figure 1 Further extensions and variations are described in the context of Figure 5 or elsewhere in this document.

[0083] Example method extension

[0084] Figures 5A to 5B The flowcharts 501-502 are examples of extensions to the first method.

[0085] Figure 5A The extension continues from box 430 and illustrates the use of two scans to determine the orientation of the interface between two materials. In the example, the orientation can be obtained by scanning along a first predetermined path. Figure 4The image is acquired at frame 410. At process frame 510, the actions of frames 410 (acquisition) and 420 (classification) can be repeated for a second set of images acquired on a second predetermined path different from the first path. Therefore, at frame 520, the second boundary position of the first material can be determined along the second path. Then at frame 530, the orientation of the interface of the first material can be determined using the two boundary positions determined at frames 430 and 520. For example, a straight line (or the normal to the straight line) connecting the first boundary position and the second boundary position can indicate the interface orientation.

[0086] In an alternative example, further scans can be performed to identify further boundary locations, and in the case of multiple such boundary locations, the boundaries of the first material can be traced for any boundary shape.

[0087] Figure 5B The extension continues from box 430 and illustrates the application of the disclosed techniques in an automated workflow. As an example, the workflow objective could be to acquire a TEM image at a precise location (e.g., offset by 200 nm) relative to the top surface of the bulk single-crystal material. In this workflow, Figure 4 This method can be used to locate the boundary between single-crystal and polycrystalline materials using a first imaging mode (e.g., SEM, STEM, or CBED). Subsequently, a TEM beam can be precisely positioned at a desired offset from the boundary, and the desired TEM image can be acquired. In variations, the workflow objective may be to perform FIB milling or electrical probing at a precise location relative to the boundary.

[0088] At box 540, a second imaging mode, different from the first imaging mode, can be used to acquire another image of the sample at a coordinate position that has a predetermined spatial relationship with the boundary position defined at box 430. In this example, the second imaging mode may be a transmission electron microscope.

[0089] As a further extension, it can be achieved through repetition. Figure 4 or Figure 5A The method determines multiple boundary locations or boundaries, and the coordinate position for performing a second imaging mode can be set relative to these multiple boundary locations or boundaries. In some examples, a common beamline can be used for both imaging modes.

[0090] Example Neural Network

[0091] Figure 6 This is a diagram 600 illustrating an example neural network that can be used to implement the disclosed techniques. As shown, neural network 620 in... Figure 6 The upper left corner receives the reciprocal space representation of the imaging region, and the lower right corner outputs the classification of materials in the imaging region.

[0092] The description begins with the input image. In a variation of the example, the disclosed technique can be implemented using an imager that generates an image of reciprocal space or physical space. In the former case, the reciprocal image 612 can be directly input into the neural network 620, as shown. In the latter case, the direct image 614 can be transformed (e.g., by FFT) at process block 616 to generate a reciprocal space representation of the imaging region, which can then be input into the neural network 620, as shown.

[0093] The neural network 620 is illustrated as a residual neural network (“ResNet”), which is a deep neural network with multiple convolutional layers having bypass skip connections. ResNet can be constructed to reshape a wide, shallow representation of the input data space into a narrow, deep representation of the output data space. As data flows through the neural network, the data space can evolve from an image representation to a classification representation. However, for ease of description, this paper describes the entire data array of the neural network as an array of pixels.

[0094] In the example, the input data space can be a reciprocal space image of the sample region, similar to... Figure 2 One of images 202-205. For illustration, the input data size can be a 1024×1024 pixel intensity array, and since the image data is monochrome, there can be only one slice of this size. Therefore, the input data size can be 1024×1024×1. In a variation, RGB color data with three slices can be used, in which case the input data size can be 1024×1024×3. These image array sizes are exemplary. Slice sizes from 64×64 to 4096×4096 are common, and other smaller, larger, or different shaped slices can also be used. Specifically, the slice sizes do not need to be equal or powers of 2.

[0095] A. Convolutional block 650

[0096] As shown in the figure, the neural network 620 contains multiple instances of convolutional blocks 650 labeled 650A-650J, as well as several other layers. Figure 6 In the upper right corner, block 650 is shown in parametric form, having J input slices (commonly referred to in the art as a "batch"), K output slices, N×N convolutions, and downsampling by a factor S. Block 650 receives J slices of the data array, copies of which are processed in parallel along two paths.

[0097] The left side of block 650 depicts the convolution path. At block 651, N×N convolution operators are applied across each slice of the input data array, and up to K×J such N×N matrices can exist to cover all combinations of J input slices and K output slices. Additionally, ÷S downsampling can be performed at block 651. For example, ÷2 downsampling can reduce a 1024×1024 data slice to 512×512, which compensates for the increase in the number of slices as the data flows down the neural network 620. For S=1, no downsampling is performed, and a given slice of input data retains its size after processing. Then, at block 652, batch normalization can be performed across the K slices output from block 651, followed by a rectified linear unit layer (RelU) at block 653, which introduces nonlinearity into the transfer function of block 650 or neural network 620. The output of block 653 can be fed into another set of N×N convolutional filters (which can be K×k) at block 654, followed by another batch of normalization stages 655. For input data organized into J slices (each slice having R rows and C columns), the output of batch block 655 can be K slices of an array of (R / S)×(C / S).

[0098] The right side of block 650 depicts the skip path that bypasses the convolution path. At block 656, the input data array slices are downsampled by ÷S so that each array slice received from both paths is the same size at block 657. The data arrays from the two paths are then summed at block 657, and the sum is passed through another RelU layer at block 658.

[0099] B. Neural Network 620

[0100] Returning to neural network 620, the first block 650A is an instance of convolutional block 650 with J = 1 input slices, 7×7 convolutions, ÷2 downsampling, and K = 64 output slices. The next block 622 is a pooling layer that performs ÷2 downsampling. This pooling layer uses a max pooling function, such that each output pixel has a maximum value in its corresponding four-pixel (2×2) group in the input array. Following block 622 are two identical instances 650B-650C of block 650, each performing 3×3 convolutions without downsampling, with 64 input slices and 64 output slices. The next two blocks 650D-650E are instances of block 650 that perform 3×3 convolutions and generate 128 output slices. Block 650D performs ÷2 downsampling to compensate for the increased slice depth, while block 650E maintains the slice size and number of slices. Another pair of blocks, 650F-650G, are also instances of block 650, performing 3×3 convolutions and generating 256 output slices. Block 650F also performs ÷2 downsampling to compensate for the increased slice depth, while block 650G maintains the slice size and number of slices. The final instance of convolutional blocks 650 in neural network 620 is blocks 650H-650J, which perform 3×3 convolutions and generate 512 output slices. Block 650H performs ÷2 downsampling to compensate for the increased slice depth, while block 650J maintains the slice size and number of slices. Following block 650J is another ÷2 pooling stage 624. This pooling stage uses an average pooling function such that the value of each output pixel is the arithmetic mean of the four input pixels that provided the data.

[0101] By design, data shaping from block 650A to block 624 can moderately reduce the number of degrees of freedom or the total amount of data flowing through the data array of neural network 620. In this example, after six ÷ 2 downsampling stages and an increase from 1 slice to 512 slices, the total amount of data is reduced by an eightfold.

[0102] The output of block 624 provides data to the fully connected layer at block 626, which linearly combines its input with the corresponding weights. Activation block 628 operates on the output of block 626, applying a non-linear transformation to generate probability vector 632. Specifically, block 628 can constrain its output values ​​to a range [0,1] suitable for the probabilities. For a binary classifier, block 628 can be implemented using the sigmoid function, while for three or more classifiers, the softmax function can be used. Probability vector 632 can be normalized such that the sum of its elements equals one.

[0103] C. Example Classifier

[0104] In the first example, a binary classifier for detecting monocrystalline silicon outputs a two-element probability vector 632, where the first element indicates the probability of "yes" (monocrystalline Si) and the second element indicates the probability of "no" (non-monocrystalline Si). Another binary classifier can distinguish between monocrystalline and polycrystalline Si, having only two output elements, P(monocrystalline) and P(polycrystalline), where P(monocrystalline) + P(polycrystalline) = 1. A third classifier can also distinguish between monocrystalline and polycrystalline Si, but can support a third result of "neither," such that P(monocrystalline) + P(polycrystalline) + P(neither) = 1. A fourth classifier can provide additional outputs, such as probabilities for monocrystalline silicon, polycrystalline silicon, amorphous silicon, high-Z material, low-Z material, or "other" in any combination.

[0105] The probability vector 632 can be analyzed to determine which index position has the maximum value (highest probability), and this value can be mapped to the classification 634 and output, thus completing the beam image processing.

[0106] Training similar to approximately 2000 CBED training images Figure 6 An exemplary neural network is used to perform binary classification between crystalline silicon and amorphous silicon.

[0107] Example Scan

[0108] Figure 7 Figure 700 illustrates an example of using multiple scans to locate a material boundary and determine its orientation. The illustrated examples include cases where the boundary extends in a horizontal, vertical, or inclined direction (as in...). Figure 7 (as seen in the plane), and cases involving tracing the boundaries of complex shapes.

[0109] Figure 7 A cross-sectional view of sample 710 is shown, in which boundary 712 separates regions 714 and 716 with different materials A and B, which can be distinguished based on the reciprocal space representation of the image data.

[0110] Initially, regions 721-724 form a quadrilateral, which defines four paths 725-728, along which scans can be performed. A first scan can be performed along path 725 to obtain results similar to... Figure 4A series of images within frame 410. In some examples, images can be acquired step-by-step. For illustration, with regions 721 and 722 separated by a distance of 2 μm and a desired accuracy of 100 nm for determining the boundary location, path 725 can be divided into 10 intervals, each 200 nm. A series of images can be acquired along path 725 from region 721 to region 722 in the order of 0, 200, 400, 600, ... nm. The reciprocal space images can be sorted upon acquisition, or image processing can be postponed until after the scan along path 725 is complete.

[0111] In other examples, a binary search can be performed for greater efficiency. For illustration, reciprocal space images can be obtained at regions 721 and 722, and these reciprocal space images can be classified. When it is determined that regions 721 and 722 have different materials and therefore the boundary lies between regions 721 and 722, a third reciprocal space image can be obtained in the middle between regions 721 and 722 to narrow the interval where the boundary can be found. For example, in... Figure 3 As described in the context, the interval can be iteratively further subdivided until the coordinate position 761 of the boundary 712 along the path 725 has been determined to the required accuracy.

[0112] Paths 726-728 can be scanned in a similar manner. For path 726, the classification of the reciprocal images obtained at regions 722 and 723 indicates that these two regions have the same material B. Therefore, it can be determined that boundary 712 does not intersect with path 726, and the scan of path 726 can terminate after only these two image points. Similarly, it can be found that path 728 does not have a boundary intersection, while path 727 does have a boundary intersection 762, which can be found in a similar manner to path 725.

[0113] In this way, scanning along paths 725 and 727 can identify horizontally oriented boundaries. Furthermore, when boundary 712 has a rotation relative to the horizontal direction, coordinates 761 and 762 can be used to establish the orientation of the boundary.

[0114] The description then shifts to the rectangle formed by regions 731-734, which defines paths 735-738. Paths 735 and 736 are similar to paths 725 and 726; their corresponding scans result in the boundary 712's coordinate position 763 intersecting with path 735, and there is no boundary intersection along path 736. It can be observed that the scan of path 737 has the same material B at both endpoints 733 and 734 and terminates immediately, similar to paths 726 and 736. Conversely, the material classification at endpoints 731 and 734 reveals different materials, and the binary search along path 738 establishes the boundary 712's coordinate position 764.

[0115] In this way, scanning along paths 735 and 738 can identify the boundaries of the tilted orientation. Furthermore, coordinates 763 and 764 can be used to establish the orientation of the boundary 712 that traverses these paths.

[0116] The description then shifts to the square formed by regions 741-744, which defines paths 745-748. Paths 748 and 747 are similar to paths 738 and 737; their corresponding scans result in the boundary 712's coordinate position 766 intersecting with path 748, while there is no boundary intersection along path 747. It can be observed that the scan of path 745 has the same material A at both endpoints 741 and 742 and terminates immediately, similar to paths 726 and 747. Conversely, the material classification at endpoints 742 and 743 reveals different materials, and the binary search along path 746 establishes the boundary 712's coordinate position 765.

[0117] In this way, scanning along paths 746 and 748 can identify vertically oriented boundaries. Furthermore, coordinates 765 and 766 can be used to establish the orientation of boundary 712 that traverses these paths.

[0118] Now turn to the polygon formed by regions 751-754. The initial reciprocal image classification of vertices 751-754 may indicate that all regions 751-754 have the same material A, thus indicating that boundary 712 does not cross the polygon. That is, scan 755 can only terminate after its endpoints 751-752 have been classified, and the same applies to scans 756-758.

[0119] Many variations and extensions can be implemented within the scope of the disclosed technology. As shown in the figure, boundary 712 is not straight. As described, multiple scans can be used to trace the boundary 712 along the corresponding paths 725, 727, 735, 738, 746, 748 at coordinate positions 761-766. On the other hand, Figure 7 The scan is shown as grouping along the edges of a square (e.g., paths 725-728 form a square), but this is not required. In other examples, any closed polygon can be used. The boundary intersection along any two edges of the polygon indicates the presence and orientation of the boundary as it passes through the polygon. The reported error in orientation (angle) may be proportional to the roughness of the boundary.

[0120] Additional Examples

[0121] The following paragraphs describe additional numbered embodiments of the disclosed technology.

[0122] Example 1 is an apparatus comprising: an imager configured to acquire multiple images of a corresponding region of a sample; and a controller coupled to the imager and configured to: for each image: process a reciprocal space representation of the corresponding image; and in response to the processing: classify materials within the corresponding region among a predetermined plurality of materials; and determine the coordinate position of a boundary of a first material among the plurality of materials.

[0123] Example 2 includes the subject matter described in Example 1, and further specifies the imager as a reciprocal space imager, and each of the acquired images includes the reciprocal space representation.

[0124] Example 3 includes the subject matter described in Example 2, and further specifies the imager as a converging beam electron diffractometer (CBED).

[0125] Example 4 includes the subject matter according to any one of Examples 1 to 3, and further specifies the imager as a microscope, and the controller is further configured to transform the corresponding image into the reciprocal space representation for each image in the image.

[0126] Example 5 includes the subject matter according to any one of Examples 1 to 4, and further specifies the imager as a transmission electron microscope (TEM) or a scanning electron microscope (SEM).

[0127] Example 6 includes the subject matter according to any one of Examples 1 to 5, and further specifies the coordinate position as a first coordinate position, and the controller is further configured to acquire another image at a second coordinate position having a predetermined spatial relationship with the first coordinate position.

[0128] Example 7 includes the subject matter described in Example 6, and further specifies that the imager is configured to operate in a first imaging mode, and the other image is acquired using a second imaging mode different from the first mode.

[0129] Example 8 includes the subject matter according to any one of Examples 1 to 7, and further specifies that the first material is monocrystalline silicon.

[0130] Example 9 includes the subject matter according to any one of Examples 1 to 8, and further specifies that the controller is configured to classify the material among any one of three or more materials.

[0131] Example 10 includes the subject matter according to any one of Examples 1 to 9, and further specifies that the classification is a binary classification that distinguishes the first material from other materials.

[0132] Example 11 includes the subject matter according to any one of Examples 1 to 10, and further specifies that the three or more materials include monocrystalline silicon and two or more of the following materials: polycrystalline silicon, amorphous silicon, high-Z materials or low-Z materials.

[0133] Example 12 includes the subject matter according to any one of Examples 1, and further specifies that the controller includes a trained neural network configured to classify the material.

[0134] Example 13 includes the subject matter according to any one of Examples 1, and further specifies that the coordinate position is determined within a tolerance of 100 nm.

[0135] Example 14 is a method comprising: acquiring a series of images including a reciprocal space representation of a corresponding region of a sample; for each image: classifying each image to identify a corresponding material present in the corresponding region among a plurality of materials; and determining, based on the classification, the boundary location of a first material among the plurality of materials.

[0136] Example 15 includes the subject matter described in Example 14, and further specifies that the series of images is a first scan on a first predetermined path, the boundary position is a first boundary position, and the method further includes: repeating the acquisition, classification and determination action for a second scan on a second predetermined path different from the first predetermined path to identify a second boundary position of the first material.

[0137] Example 16 includes the subject matter described in Example 15, and further includes: determining the orientation of the interface of the first material based on the first boundary location and the second boundary location.

[0138] Example 17 includes the subject matter according to any one of Examples 15 to 16, and further includes: performing an additional scan; identifying additional boundary locations from the additional scan; and tracing the boundary of the first material based on the first boundary location, the second boundary location, and the additional boundary location.

[0139] Example 18 includes the subject matter of any one of Examples 14 to 17, and further specifies that the series of images includes three or more images; and wherein the series of images is obtained by performing a binary search on a predetermined path, wherein a third image and any subsequent images of the image are obtained along the predetermined path in a corresponding region between (i) the most recent image of the image and (ii) another previous image of the image.

[0140] Example 19 includes the subject matter of any one of Examples 14 to 18, and further includes: confirming the boundary location by the following steps: acquiring two additional images in the corresponding area along the predetermined path on the opposite side of the boundary location; classifying the two additional images; and verifying that exactly one of the two additional images indicates the presence of the first material.

[0141] Example 20 includes the subject matter according to any one of Examples 14 to 19, and further specifies that the classification is based on the detection of the following: the presence or absence of a moiré pattern, or one or more Kikuchi lines; the intensity of the central spot or halo; and the symmetry of one or more features of each image.

[0142] Example 21 includes the subject matter according to any one of Examples 14 to 20, and further refers to performing the classification independently for each image in the image.

[0143] Example 22 includes the subject matter according to any one of Examples 14 to 21, and further specifies that the sample includes a manufactured semiconductor device.

[0144] General Computer Environment

[0145] Figure 8 General examples of suitable computing systems 800 are illustrated, in which the described examples, techniques, and methods can be implemented to apply focus stacking to sample preparation or analysis processes. The computing system 800 is not intended to impose any limitation on the scope or functionality of this disclosure, as these innovations can be implemented in a variety of general-purpose or special-purpose computing systems. The computing system 800 can control a CBED imager, FIB milling machine, microscope, stage, analyzer, or other similar equipment; can perform or control image acquisition, image transformation, scanning of sample surfaces, training or application of neural networks, material classification, metrology, or other analysis of images or other acquired data representing the sample; can control a stage, ion beam column, or electron beam column; or can acquire, process, output, or store measurement data.

[0146] refer to Figure 8 The computing environment 810 includes one or more processing units 822 and a memory 824. Figure 8In this document, the basic configuration 820 is included within the dashed lines. Processing unit 822 can execute computer-executable instructions, such as those for control, metering, or other functions as described herein. Processing unit 822 can be a general-purpose central processing unit (CPU), a processor in an application-specific integrated circuit (ASIC), or any other type of processor. In a multiprocessor system, multiple processing units execute computer-executable instructions to enhance processing power. Computing environment 810 may also include a graphics processing unit or a coprocessor unit 830. Tangible memory 824 can be volatile memory (e.g., registers, cache, or RAM), non-volatile memory (e.g., ROM, EEPROM, or flash memory), or a combination thereof, accessible by processing units 822 and 830. Memory 824 stores software 880 implementing one or more innovations described herein in a form suitable for execution by computer-executable instructions from processing units 822 and 830. For example, software 880 may include software 881 for controlling a CBED, SEM, or other imager; software 882 for controlling a FIB or other milling tool; software 883 for controlling a stage supporting a sample thereon; software 884 for implementing a trained neural network or other classifier; software 885 for performing metrology or other analysis on the sample data; or other software 886 (including a user interface, host interface, fault detection, or neural network training). The illustrations shown for software 880 in storage device 840 are similarly applicable to... Figure 8 The software 880 is located elsewhere in the memory. The memory 824 may also store control parameters, calibration data, measurement data, other database data, configuration data, or operation data.

[0147] The computing system 810 may have additional features such as one or more of a storage device 840, an input device 850, an output device 860, or a communication port 870. Interconnection mechanisms (not shown), such as buses, controllers, or networks, interconnect the components of the computing environment 810. Typically, operating system software (not shown) provides an operating environment for other software 880 executing in the computing environment 810 and coordinates the activities of the components of the computing environment 810.

[0148] The physical storage device 840 may be removable or non-removable and includes a magnetic disk, magnetic tape or cassette tape, CD-ROM, DVD, or any other medium that can be used to store information in a non-transitory manner and is accessible within the computing environment 810. The storage device 840 stores instructions (including instructions and / or data) for implementing one or more innovative software 880 described herein. The storage device 840 may also store image data, measurement data, workflow sequences, reference data, calibration data, configuration data, sample data, or other databases or data structures described herein.

[0149] Input device 850 may be a mechanical, touch-sensing, or proximity-sensing input device (such as a keyboard, mouse, pen, touchscreen, or trackball), a voice input device, a scanning device, or another device that provides input to computing environment 810. Output device 860 may be a display, printer, speaker, optical disc writer, or another device that provides output from computing environment 810. Input or output may also be transmitted to / from a remote device via a network connection through communication port 870 (e.g., ...). Figure 1 (as described in the context).

[0150] Communication port 870 enables communication with another computing entity via a communication medium. The communication medium transmits information such as computer-executable instructions, audio or video input or output, or other data in a modulated data signal. A modulated data signal is a signal having one or more of its characteristics set or altered in a manner that encodes information in the signal. By way of example and not limitation, the communication medium may be electrical, optical, RF, acoustic, or other carriers.

[0151] The data acquisition system can be integrated into the computing environment 810 as an input device 850 or coupled to the communication port 870, and may include an analog-to-digital converter or a connection to an instrument bus. The instrument control system can be integrated into the computing environment 810 as an output device 860 or coupled to the communication port 870, and may include a digital-to-analog converter, a switch, or a connection to an instrument bus.

[0152] In some examples, computer system 800 may also include instruction cloud 890 in which all or part of the disclosed technology is executed. Any combination of memory 824, storage device 840, and computing cloud 890 may be used to store software instructions and data of the disclosed technology.

[0153] This invention is described in the general context of computer-executable instructions (e.g., those contained in a program module) that can be executed on a computing system on a target real or virtual processor. Typically, a program module or component includes routines, programs, libraries, objects, classes, components, data structures, etc., which perform a specific task or implement a specific data type. In various implementations, the functionality of the program module can be combined or separated as needed. The computer-executable instructions for the program module can execute within a local or distributed computing system.

[0154] The terms “computing system,” “computing environment,” and “computing device” are used interchangeably throughout this document. Unless the context explicitly indicates otherwise, no term implies any limitation on the type of computing system, computing environment, or computing device. Generally, a computing system, computing environment, or computing device can be local or distributed and can contain any combination of dedicated hardware and / or general-purpose hardware and / or virtualization hardware, as well as software that implements the functionality described herein. Virtual processors, virtual hardware, and virtualization devices are ultimately embodied in hardware processors or another form of physical computer hardware, and therefore include both the software and underlying hardware associated with virtualization.

[0155] General considerations

[0156] As used herein and in the claims, the singular forms “a,” “an,” and “the” include the plural forms unless the context clearly indicates otherwise. Additionally, the term “comprising” means “including.” Furthermore, the term “coupled” does not exclude the existence of intermediate elements between coupled items. Moreover, as used herein, the terms “or” and “and / or” mean any one or more of the phrases.

[0157] The systems, methods, and apparatuses described herein should not be construed as limiting in any way. Rather, this disclosure relates to all novel and non-obvious features and aspects of the various disclosed embodiments, whether individually or in various combinations and sub-combinations formed with each other. The disclosed systems, methods, and apparatuses are not limited to any particular aspect or feature or combination thereof, nor are they required to have any one or more particular advantages or problems solved. Techniques from any example may be combined with one or more of the techniques described in any other example.

[0158] Although some operations in the disclosed methods are described in a specific sequential order for ease of presentation, it should be understood that this descriptive approach encompasses the rearrangement of operations unless the specific language used herein requires a particular order. For example, operations described sequentially may be rearranged or performed simultaneously in some cases. Furthermore, for simplicity, the accompanying drawings may not show the various ways in which the disclosed systems, methods, and apparatus can be combined with other systems, methods, and apparatuses. Additionally, the description sometimes uses terms such as “acquire,” “activate,” “apply,” “average,” “divide,” “classify,” “configure,” “control,” “convolution,” “determine,” “downsampling,” “filter,” “generate,” “recognize,” “image,” “pool,” “process,” “produce,” “scan,” “select,” “search,” “set,” “train,” “transform,” or “verify” to describe the disclosed methods. These terms are high-level abstractions of the actual operations performed. The actual operations corresponding to these terms will vary depending on the specific implementation and can be readily identified by one of skill in the art.

[0159] In some examples, values, programs, or devices are referred to as “lowest,” “best,” “maximum,” “optimal,” “extreme,” etc. It will be understood that such descriptions are intended to indicate that a choice can be made among several or many alternatives, and that such a choice is not necessarily better, smaller, or otherwise preferred than other unconsidered alternatives.

[0160] The operational theories, scientific principles, or other theoretical descriptions presented herein with reference to the apparatus or methods of this disclosure are provided for the purpose of better understanding and interpretation, and are not intended to be limiting in scope. That is, the disclosed systems, methods, and apparatus are not limited to such operational theories. The appended claims are not limited to embodiments that operate in a manner described by such operational theories.

[0161] Any of the disclosed methods can be controlled or implemented as a computer-executable instruction or computer program product stored on one or more computer-readable storage media (e.g., tangible, non-transitory computer-readable storage media) and executed on a computing device (e.g., any available computing device, including tablets, smartphones, or other mobile devices containing computing hardware). A tangible computer-readable storage medium is any available tangible medium that can be accessed in a computing environment (e.g., one or more optical media such as a DVD or CD, a volatile memory component (e.g., DRAM or SRAM), or a non-volatile memory component (e.g., flash memory or hard disk drive)). By way of example and reference Figure 8The computer-readable storage medium includes memory 824 and storage device 840. The terms "computer-readable medium" or "computer-readable storage medium" do not include signals and carrier waves. Additionally, the terms "computer-readable medium" or "computer-readable storage medium" do not include communication ports (e.g., 870).

[0162] Any of the computer-executable instructions used to implement the disclosed technology, and any data created and used during the implementation of the disclosed embodiments, may be stored on one or more computer-readable storage media. The computer-executable instructions may be, for example, part of a dedicated software application or a software application accessed or downloaded via a web browser or other software application (e.g., a remote computing application). Such software may be executed using one or more networked computers, for example on a single local computer (e.g., any suitable commercially available computer) or in a networked environment (e.g., via the Internet, a wide area network, a local area network, a client-server network, a cloud computing network, or other such networks).

[0163] For clarity, only selected aspects of software-based implementations are described. Other details well-known in the art are omitted. For example, it should be understood that the disclosed technology is not limited to any particular computer language or program. For instance, the disclosed technology can be implemented in software written in Adobe Flash, C, C++, C#, Curl, Dart, Fortran, Java, JavaScript, Julia, Lisp, Matlab, Octave, Perl, Python, Qt, R, Ruby, SAS, SPSS, SQL, WebAssembly, any derivative thereof, or any other suitable programming language, or in some examples, markup languages ​​such as HTML or XML, or any combination of suitable languages, libraries, and data packages. Similarly, the disclosed technology is not limited to any particular type of computer or hardware. Certain details of suitable computers and hardware are well-known and do not need to be elaborated in this disclosure.

[0164] Furthermore, any of the software-based implementation schemes (including, for example, computer-executable instructions for causing a computer to perform any of the disclosed methods) can be uploaded, downloaded, or remotely accessed via suitable communication means. Such suitable communication means include, for example, the Internet, the World Wide Web, corporate intranets, software applications, cable (including fiber optic cables), magnetic communication, electromagnetic communication (including RF, microwave, infrared, and optical communication), electronic communication, or other such communication means.

[0165] Given the many possible embodiments to which the principles of the disclosed subject matter can be applied, it should be recognized that the illustrated embodiments are merely preferred examples and should not be considered as limiting the scope of the claims. Rather, the scope of the claimed subject matter is defined by the appended claims. Therefore, we claim protection for all contents falling within the substance and spirit of these claims.

Claims

1. An apparatus for sample analysis, the apparatus comprising: An imager configured to acquire multiple images of a corresponding region of a sample; and A controller, coupled to the imager and configured to: For each of the images: Process the reciprocal space representation of the image; as well as In response to the processing, the materials within the corresponding region are classified among any one of three or more predetermined materials, wherein the three or more materials include monocrystalline silicon and two or more of the following: polycrystalline silicon, amorphous silicon, high-Z materials, or low-Z materials; and Determine the coordinate position of the boundary of the first material among the multiple materials.

2. The device of claim 1, wherein the imager is a reciprocal space imager and each of the acquired images includes the reciprocal space representation.

3. The device according to claim 2, wherein the imager is a converging beam electron diffractometer (CBED).

4. The device of claim 1, wherein the imager is a microscope, and the controller is further configured to transform the corresponding image into the reciprocal space representation for each image in the images.

5. The device of claim 1, wherein the imager is a transmission electron microscope (TEM).

6. The device of claim 1, wherein the coordinate position is a first coordinate position, and the controller is further configured to acquire another image at a second coordinate position having a predetermined spatial relationship with the first coordinate position.

7. The device of claim 6, wherein the imager is configured to operate in a first imaging mode, and the other image is acquired using a second imaging mode different from the first imaging mode.

8. The device of claim 1, wherein the controller includes a trained neural network configured to classify the material.

9. The device according to claim 1, wherein the coordinate position is determined within a tolerance of 100 nm.

10. A method for analyzing a sample, the method comprising: A series of images are acquired as a first scan along a first predetermined path, the images including the reciprocal space representation of the corresponding regions of the sample; For each of the images: Each image is classified to identify the corresponding material present in the corresponding region among multiple materials; Based on the classification, the first boundary position of the first material among the multiple materials is determined; as well as The acquisition, classification, and determination actions are repeated for a second scan on a second predetermined path different from the first predetermined path to identify the second boundary location of the first material.

11. The method according to claim 10, further comprising: The orientation of the interface of the first material is determined based on the first boundary position and the second boundary position.

12. The method according to claim 10, further comprising: Perform an additional scan; Identify additional boundary locations from the additional scan; The boundaries of the first material are traced based on the first boundary location, the second boundary location, and the additional boundary location.

13. The method of claim 10, wherein the classification of each image of the first scan is based on detecting the presence or absence of: Moiré pattern, or One or more Kikuchi lines.

14. The method of claim 10, wherein the classification of each of the images in the first scan is performed independently for each of the images in the first scan.

15. The method of claim 10, wherein the sample comprises a manufactured semiconductor device.

16. The method of claim 10, wherein the classification of each image of the first scan is based on the intensity of the detected central spot or halo.

17. The method of claim 10, wherein the classification of each image of the first scan is based on the symmetry of one or more features of each image detected.

18. A method for sample analysis, comprising: A series of three or more images are obtained by performing a binary search along a predetermined path, the three or more images comprising the reciprocal space representation of the corresponding regions of the sample; Wherein, a third image and any subsequent image of the three or more images are obtained at a corresponding region located along the predetermined path between (i) the most recent image of the image and (ii) another previous image of the image; For each of the three or more images: Each image is classified to identify the corresponding material present in the corresponding region among multiple materials; as well as Based on the classification, the boundary position of the first material among the multiple materials is determined.

19. The method of claim 18, further comprising: The boundary location is confirmed through the following steps: Two additional images are acquired in the corresponding regions along the predetermined path on the opposite side of the boundary location; as well as Classify the two additional images; and Verify that exactly one of the two additional images indicates the presence of the first material.

20. The method of claim 18, wherein the classification is based on detecting the following: The presence or absence of a Mohr pattern, or one or more Kikuchi lines; The intensity of the central spot or halo; or The symmetry of one or more features of each image.

21. The method of claim 18, wherein the classification is performed independently for each of the three or more images.

22. The method of claim 18, wherein the sample comprises a manufactured semiconductor device.

Images