Training a machine learning model to produce higher resolution images from inspection images

Abstract

Methods and systems for determining information of a specimen are provided. Embodiments described herein are configured to train a machine learning (ML) model for generating a higher resolution image of a specimen from an image of the specimen generated by an inspection subsystem. The training includes a pre-training step performed using only simulated images and a re-training step performed using actual images of test specimens. The higher resolution images generated from lower resolution inspection images by the trained ML model can be used for applications including nuisance filtering and defect classification.

Description

Technical Field

[0001] This invention generally relates to methods and systems for determining information from samples using a trained machine learning model. Background Technology

[0002] The following descriptions and examples are not considered prior art because they are included in this section.

[0003] Semiconductor devices, such as logic and memory devices, typically involve processing a substrate, such as a semiconductor wafer, using numerous semiconductor processes to form various features and multiple layers of the semiconductor device. For example, photolithography is a semiconductor process involving transferring a pattern from a photomask to a photoresist disposed on a semiconductor wafer. Additional examples of semiconductor processes include (but are not limited to) chemical mechanical polishing (CMP), etching, deposition, and ion implantation. Multiple semiconductor devices can be fabricated on a single semiconductor wafer and then separated into individual semiconductor devices.

[0004] Inspection processes are used at various steps during semiconductor manufacturing to detect defects in samples, driving higher yields and thus higher profits. Inspection has always been a crucial part of manufacturing semiconductor devices. However, as the size of semiconductor devices shrinks, inspection becomes even more critical for successfully manufacturing acceptable semiconductor devices, as even small defects can cause device failure.

[0005] Defect inspection typically involves re-inspecting defects, such as those detected by the inspection process, using high-magnification optical systems or scanning electron microscopy (SEM) to generate additional information about the defects at higher resolution. Therefore, defect inspection is performed at discrete locations on the sample where defects were detected by inspection. The higher-resolution data for defects generated by defect inspection is better suited for determining defect properties (e.g., profile, roughness, more accurate size information, etc.).

[0006] Measurement processes are also used to monitor and control the process at various stages of semiconductor manufacturing. Measurement processes differ from inspection processes in that, unlike inspection processes which detect defects on samples, measurement processes measure one or more characteristics of the sample that cannot be determined using currently available inspection tools. For example, measurement processes measure one or more characteristics of the sample, such as the dimensions of features formed on the sample during the process (e.g., linewidth, thickness, etc.), allowing the performance of the process to be determined from one or more characteristics. Furthermore, if one or more characteristics of the sample are unacceptable (e.g., outside a predetermined range), the measurement of these characteristics can be used to change one or more parameters of the process, resulting in additional samples manufactured by the process having acceptable characteristics.

[0007] The measurement process also differs from the defect inspection process in that, unlike defect inspection, which involves revisiting defects detected by the inspection institute, the measurement process can be performed at locations where no defects were detected. In other words, unlike defect inspection, the location for performing a measurement process on a sample can be independent of the results of the inspection process performed on the sample. Specifically, the location for performing the measurement process can be chosen independently of the inspection results.

[0008] As described above, due to the limited resolution of the inspection process (optical and sometimes electron beam inspection), additional higher-resolution images of the sample are generally required for defect inspection, which may include one or more of the following: verifying defect detection, classifying defect detection, and determining the characteristics of the defect. Furthermore, higher-resolution images are generally needed to determine information about patterned features formed on the sample, as in metrology, regardless of whether a defect has been detected in the patterned features. Therefore, defect inspection and measurement can be a time-consuming process requiring the physical sample itself and the tools needed to generate higher-resolution images.

[0009] However, the process of defect inspection and measurement cannot be simply eliminated to save time and money. For example, due to the resolution of the inspection process, it generally does not generate image signals or data that could be used to identify defects sufficient to classify them and / or determine their root causes. Furthermore, due to the resolution of the inspection process, it generally does not generate image signals or data that could be used to determine patterned features formed on a sample with sufficient accuracy and / or precision.

[0010] There are two main categories of methods previously used to improve the resolution of optical images via computer-based post-processing. The first category is super-resolution methods that use multiple views from the same location. Here, multiple optical modes, focus shifts, and / or perspectives are used to acquire a lower-resolution image. The higher-resolution image is reconstructed using an optical model or a learning method with a deep learning (DL) model. The second category is DL models that reconstruct higher-resolution images using only the optical image. This relies on learning the basic distribution of features of the higher-resolution image, such as texture, edges, and object shape models.

[0011] However, current computer-based methods for improving optical image resolution have several drawbacks. For example, super-resolution methods significantly impact the output of optical tools. The slowdown factor is proportional to the number of views required for high-resolution reconstruction. When the tool typically samples at its Nyquist rate, model-based methods are prone to artifacts due to limited information in the optical image.

[0012] Therefore, it would be advantageous to develop systems and methods for determining information about samples that do not have one or more of the aforementioned drawbacks. Summary of the Invention

[0013] The following description of various embodiments is not to be construed in any way as limiting the subject matter of the appended claims.

[0014] One embodiment relates to a system configured to determine information about a sample. The system includes a verification subsystem configured to generate an image of the sample. The system also includes one or more computer subsystems and one or more components executed by the computer subsystems. The one or more components include a trained machine learning (ML) model configured to transform the image of the sample generated by the verification subsystem into a higher resolution image of the sample.

[0015] The one or more computer subsystems are configured to pretrain an initial ML model using a pretraining set, thereby generating a pretrained ML model. The pretraining set includes simulated test images of test samples designated as pretraining inputs and corresponding simulated higher-resolution images of the test samples designated as pretraining outputs.

[0016] The one or more computer subsystems are also configured to retrain the pre-trained ML model using a training set, thereby generating the trained ML model. The training set includes images generated by the verification subsystem for the test samples, designated as training inputs, and corresponding higher-resolution images of the test samples generated by the high-resolution imaging system, designated as training outputs.

[0017] The one or more computer subsystems are further configured to input the image generated by the testing subsystem for the sample during testing into the trained ML model to generate the higher resolution image of the sample. Additionally, the one or more computer subsystems are configured to determine information about the sample from the generated higher resolution image. The system may be further configured as described herein.

[0018] Another embodiment relates to a computer-implemented method for determining information about a sample. The method includes the pre-training, retraining, input, and determination steps described above, performed by one or more computer systems. One or more components are executed by the one or more computer systems. The one or more components include the trained ML model, which is configured to transform images of the samples generated by the verification subsystem into higher-resolution images of the samples.

[0019] Each of the steps of the method may be performed as further described herein. The method may include any other steps of any other method described herein. The method may be performed by any of the systems described herein.

[0020] Another embodiment relates to a non-transitory computer-readable medium storing program instructions that can be executed on one or more computer systems for determining information of a sample. The computer-implemented method includes the steps of the methods described above. The computer-readable medium may be further configured as described herein. The steps of the computer-implemented method may be performed as further described herein. Additionally, the computer-implemented method to which the program instructions are executable may include any other steps of any other method described herein. Attached Figure Description

[0021] Those skilled in the art will appreciate further advantages of the invention from the following detailed description of preferred embodiments and from reference to the accompanying drawings, wherein:

[0022] Figure 1 and 1a This is a schematic diagram of a side view illustrating an embodiment of a system configured as described herein;

[0023] Figure 2 It is a plan view illustrating the location of defects of interest (DOI) and disturbance points in the design of the sample, a low-resolution image of DOI and disturbance points on the sample, and a schematic diagram of a high-resolution image of DOI and disturbance points on the sample.

[0024] Figure 3 This is a flowchart illustrating an embodiment of steps that can be performed to pre-train an initial machine learning (ML) model;

[0025] Figure 4 This is a flowchart illustrating an embodiment of steps that can be executed to generate simulated test images and simulated higher resolution images of a pre-training set;

[0026] Figure 5 This is a flowchart illustrating an embodiment of steps that can be performed to retrain a pre-trained ML model;

[0027] Figure 6 This is a flowchart illustrating an embodiment of a trained ML model during runtime.

[0028] Figure 7 This is a flowchart illustrating an embodiment of steps that can be performed to determine information about a sample; and

[0029] Figure 8This is a block diagram illustrating an embodiment of a non-transitory computer-readable medium storing program instructions for causing a computer system to perform the computer-implemented methods described herein.

[0030] While various modifications and alternatives are permissible with respect to the invention, specific embodiments thereof are shown by way of example in the drawings and described in detail herein. The drawings may not be drawn to scale. However, it should be understood that the drawings and their detailed description are not intended to limit the invention to the specific forms disclosed, but rather are intended to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims. Detailed Implementation

[0031] As used interchangeably herein, the terms “design,” “design data,” and “design information” generally refer to the physical design (layout) of an IC or other semiconductor device and the data derived from the physical design through complex simulations or simple geometric and Boolean operations. A design may include any other design data or design data proxies described in U.S. Patent No. 7,570,796, jointly owned by Zafar et al., issued August 4, 2009, and U.S. Patent No. 7,676,077, jointly owned by Kulkarni et al., issued March 9, 2010, which are incorporated herein by reference as if fully set forth herein. Additionally, design data may be standard cell library data, integrated layout data, design data for one or more layers, derivatives of design data, and all or part of chip design data. Furthermore, the terms “design,” “design data,” and “design information” as used herein refer to information and data generated by a semiconductor device designer during the design process and are therefore available for use in the embodiments described herein prior to printing the design onto any physical sample (e.g., a photomask and a wafer).

[0032] As used in this article, the term "design clip" is generally defined as a relatively small portion of the overall design of a sample. The term "design clip" is used interchangeably with the phrase "a portion of the design data" in this article.

[0033] As used herein, the term "disruptive point" (which may sometimes be used interchangeably with "disruptive point defect" or "disruptive point event") is generally defined as a defect that is of no concern to the user and / or an event detected on a sample that is not actually a defect on the sample. Disruptive points that are not actual defects can be detected as events attributable to non-defect noise sources on the sample (e.g., grains in the metal lines on the sample, signals from the underlying material or material on the sample, line edge roughness (LER), relatively small critical dimension (CD) variations in patterned features, thickness variations, etc.) and / or to the inspection system itself or edges in its configuration for inspection.

[0034] As used herein, the term "DOI" is defined as a defect that is detected on a sample and is, in fact, an actual defect on the sample. Therefore, a DOI is of user interest because the user is generally concerned with how many and what kinds of actual defects are present on the sample being examined. In some contexts, the term "DOI" is used to refer to a subset of all actual defects on a sample, containing only those actual defects of interest to the user. For example, multiple types of DOIs may exist on any given sample, and the user may be more concerned with one or more of these DOIs than with one or more other types. However, in the context of the embodiments described herein, the term "DOI" is used to refer to any and all true defects on a sample.

[0035] As used herein, the term "low-resolution image" generally refers to an image in which all patterned features formed in the sample region that produces the image are not resolved in the image. For example, if some patterned features in the sample region that produces the low-resolution image are large enough to be resolved, then they can be resolved in the low-resolution image. However, a low-resolution image is not produced at a resolution that makes all patterned features in the image resolvable. In this way, as used herein, the term "low-resolution image" does not contain information about the patterned features on the sample that would be sufficient for the low-resolution image to be used for applications such as defect inspection (which may include defect classification and / or verification) and metrology. Additionally, as used herein, the term "low-resolution image" generally refers to an image produced by an inspection system that typically has a relatively low resolution (e.g., lower than that of a defect inspection and / or metrology system) for relatively fast output.

[0036] "Low-resolution image" can also mean "low-resolution" because it has a lower resolution than the "high-resolution image" described herein. The term "high (or higher) resolution image," as used herein, can generally be defined as an image in which all patterned features of a sample are resolved with relatively high accuracy. In this way, all patterned features in the sample region for which a high-resolution image is generated are resolved in the high-resolution image, regardless of their size. Thus, the term "high-resolution image," as used herein, contains information about the patterned features of the sample sufficient to enable the high-resolution image to be used for applications such as defect inspection (which may include defect classification and / or verification) and metrology. Additionally, the term "high-resolution image," as used herein, generally refers to an image that cannot be generated by the inspection system during normal operation and is configured to sacrifice resolution capability to increase output.

[0037] Turning now to the figures, it should be noted that the figures are not drawn to scale. Specifically, some elements in the figures have been enlarged to emphasize their characteristics. It should also be noted that the figures are not drawn to the same scale. Elements that can be similarly configured and shown in more than one figure are indicated using the same element reference numerals. Unless otherwise stated herein, any element described and shown may include any suitable commercially available element.

[0038] Generally, the embodiments described herein are configured to generate higher-resolution images (e.g., optical images) using a design. The embodiments employ deep learning (DL) methods that reconstruct high-resolution optical or other images from lower-resolution images obtained from a testing tool, aided by high-resolution design information. The embodiments can leverage the availability of intended design data for two key steps. The first is a pre-trained network, which reduces the tooling time required to acquire training images because a much smaller set of non-simulated sample images is used in the final training step. Second, the design data can be used as an additional information channel for high-resolution reconstruction.

[0039] Currently, most advanced inspections (such as broadband plasma (BBP) optical inspection scanning) sample at the Nyquist rate. At the Nyquist sampling rate, image resolution is at the information theory limit of the inspection system. Any higher resolution would require changes to the light source and imaging optics. The embodiments described herein provide a computer-based method for reconstructing high-resolution images. The embodiments rely on the assumption that process variations occurring on the sample are a perturbable subset of the design that produces the observed inspection images. Analyzing the structure at higher resolution improves sensitivity and helps in accurately classifying defect types. Increasing the image resolution in post-processing can potentially improve the overall inspection tool because the inspection can be run at lower magnification.

[0040] In some embodiments, the sample is a wafer. The wafer may comprise any wafer known in semiconductor technology. Although some embodiments may be described herein with respect to one or more wafers, the embodiments are not limited to the use of their samples. For example, the embodiments described herein may be used as samples of, for example, photomasks, tablets, personal computer (PC) boards, and other semiconductor samples.

[0041] The embodiments relate to a system configured to determine information about a sample. Figure 1 An embodiment of this system is illustrated below. The system includes one or more computer subsystems 102 and one or more components 104 executed by the one or more computer subsystems. In some embodiments, the system includes a verification subsystem coupled to the one or more computer subsystems. For example, in Figure 1 The system includes a verification subsystem 100 coupled to the computer subsystem 102. Figure 1In the embodiments shown, the inspection subsystem is configured as a light-based inspection subsystem. However, in other embodiments described herein, the inspection subsystem is configured as an electron beam or charged particle beam inspection subsystem.

[0042] The testing subsystem is configured to generate an image of the sample. Generally, the testing subsystem described herein includes at least a power source, a detector, and a scanning subsystem. The power source is configured to generate energy directed to the sample by the testing subsystem. The detector is configured to detect the energy from the sample and generate an output in response to the detected energy. The scanning subsystem is configured to change the position on the sample where the energy is directed to and from which the energy is detected.

[0043] In the light-based testing subsystem, the energy directed to the sample contains light, and the energy detected from the sample also contains light. Figure 1 In the embodiment of the system shown, the testing subsystem includes an illumination subsystem configured to direct light to sample 14. The illumination subsystem includes at least one light source, such as... Figure 1 The light source 16 shown in the image. The illumination subsystem can be configured to guide light to the sample at one or more incident angles, said one or more incident angles may include one or more tilt angles and / or one or more normal angles. For example, as Figure 1 As shown, light from light source 16 is guided through optical element 18 and then through lens 20 to sample 14 at an angle of incidence. The angle of incidence can include any suitable angle of incidence, which can vary depending on, for example, the characteristics of the sample and the defects that will be detected on the sample.

[0044] The illumination subsystem can be configured to direct light to the sample at different incident angles at different times. For example, the testing subsystem can be configured to change one or more characteristics of one or more components of the illumination subsystem, so that the light can be different from... Figure 1 The incident angle shown is directed to the sample. In this example, the testing subsystem can be configured to move the light source 16, optical element 18, and lens 20 so that light is directed to the sample at different angles of incidence or at the normal (or near the normal).

[0045] In some examples, the testing subsystem can be configured to guide light to the sample simultaneously at more than one incident angle. For example, the illumination subsystem may include more than one illumination channel, one of which may include, for instance, […]. Figure 1The light source 16, optical element 18, and lens 20 shown herein, and another element in the illumination channel (not shown), may contain similar elements in different or identical configurations, or may contain at least a light source and one or more other components, such as those further described herein. If this light and another light are simultaneously directed to the sample, then one or more characteristics (e.g., wavelength, polarization, etc.) of the light directed to the sample at different incident angles may be different, such that the light obtained from illuminating the sample at different incident angles can be distinguished from each other at the detector.

[0046] In another example, the lighting subsystem may consist only of a light source (e.g. Figure 1 The light from the source 16 shown in the diagram can be separated into different optical paths (e.g., based on wavelength, polarization, etc.) by one or more optical elements (not shown) of the illumination subsystem. The light from each of these different optical paths can then be directed to the sample. Multiple illumination channels can be configured to direct light to the sample simultaneously or at different times (e.g., when different illumination channels are used to sequentially illuminate the sample). In another example, the same illumination channel can be configured to direct light with different characteristics to the sample at different times. For example, in some examples, optical element 18 can be configured as a spectral filter, and the properties of the spectral filter can be changed in various ways (e.g., by swapping the spectral filter with another), so that light of different wavelengths can be directed to the sample at different times. The illumination subsystem can have any other suitable configuration known in the art for sequentially or simultaneously directing light with different or the same characteristics to the sample at different or the same incident angles.

[0047] Light source 16 may comprise a broadband plasma (BBP) source. In this manner, the light generated by the source and directed to the sample may comprise broadband light. However, the source may comprise any other suitable source, such as a laser. The laser may comprise any suitable laser known in the art and may be configured to produce light of any suitable wavelength known in the art. The laser may be configured to produce monochromatic or near-monochromatic light. In this manner, the laser may be a narrowband laser. The source may also comprise a multicolor source that produces light of multiple discrete wavelengths or bands.

[0048] Light from optical element 18 can be focused onto sample 14 by lens 20. Although lens 20 is... Figure 1 While the lens 20 is shown as a single refractive optical element, in practice, the lens 20 may comprise a combination of several refractive and / or reflective optical elements that focus light from the optical element onto the sample. Figure 1The illumination subsystem shown and described herein may include any other suitable optical element (not shown). Examples of such optical elements include (but are not limited to) polarization components, spectral filters, spatial filters, reflective optics, apodizers, beam splitters, apertures, and similar elements to any such suitable optical element known in the art. Additionally, the system may be configured to modify one or more elements of the illumination subsystem based on the type of illumination used for testing.

[0049] The inspection subsystem also includes a scanning subsystem configured to change the position on the sample where light is directed and detected, and to cause a light scan across the sample. For example, the inspection subsystem may include a stage 22 on which sample 14 is placed during inspection. The scanning subsystem may include any suitable mechanical and / or robotic assembly (including stage 22) configured to move the sample such that light can be directed to different positions on the sample and detected from different positions on the sample. Alternatively, the inspection subsystem may be configured such that one or more optical elements of the inspection subsystem perform some scans of light across the sample, such that light can be directed to different positions on the sample and detected from different positions on the sample. In examples of causing a light scan across the sample, the light scan across the sample can be performed in any suitable manner (e.g., in a serpentine or helical path).

[0050] The inspection subsystem further includes one or more detection channels. At least one detection channel includes a detector configured to detect light originating from the sample due to illumination by the subsystem and to generate an output in response to the detection light. For example, Figure 1 The inspection subsystem shown includes two detection channels: one formed by collector 24, element 26, and detector 28, and the other formed by collector 30, element 32, and detector 34. Figure 1 As shown, two detection channels are configured to collect and detect light at different collection angles. In some examples, the two detection channels are configured to detect scattered light, and the detection channel is configured to detect light scattered from the sample at different angles. However, one or more of the detection channels may be configured to detect another type of light from the sample (e.g., reflected light).

[0051] like Figure 1As further shown, the two detection channels are positioned within the plane of the paper, and the illumination subsystem is also positioned within the plane of the paper. Therefore, in this embodiment, the two detection channels are located within the plane of incidence (e.g., at the center). However, one or more of the detection channels may be located outside the plane of incidence. For example, the detection channel formed by collector 30, element 32, and detector 34 may be configured to collect and detect light scattered out of the plane of incidence. Therefore, this detection channel may generally be referred to as a "side" channel, and this side channel may be located at the center of a plane substantially perpendicular to the plane of incidence.

[0052] although Figure 1 An embodiment of an inspection subsystem comprising two detection channels is shown, but the inspection subsystem may comprise a different number of detection channels (e.g., a single detection channel or two or more detection channels). In this example, the detection channel formed by collector 30, element 32, and detector 34 may form a side channel as described above, and the inspection subsystem may include an additional detection channel (not shown) formed as another side channel on the opposite side of the incident plane. Thus, the inspection subsystem may include a detection channel comprising collector 24, element 26, and detector 28, and the detection channel is located at the center of the incident plane and configured to collect and detect light at a scattered angle located at or nearly perpendicular to the sample surface. Therefore, this detection channel may be commonly referred to as a “top” channel, and the inspection subsystem may also include two or more side channels configured as described above. Thus, the inspection subsystem may comprise at least three channels (i.e., a top channel and two side channels), and each of the at least three channels has its own collector, each of the collectors being configured to collect light at a different scattered angle than each of the other collectors.

[0053] As further described above, each of the detection channels included in the inspection subsystem can be configured to detect scattered light. Therefore, Figure 1 The testing subsystem shown herein can be configured for dark-field (DF) testing of samples. However, the testing subsystem may also, or alternatively, include a detection channel configured for bright-field (BF) testing of samples. In other words, the testing subsystem may include at least a detection channel configured to detect light reflected from the specular surface of the sample. Therefore, the testing subsystem described herein can be configured for DF only, BF only, or both DF and BF testing. Although each of the collectors is in Figure 1 The image is shown as a single refractive optical element, but it should be understood that each of the collectors may contain one or more refractive optical elements and / or one or more reflective optical elements.

[0054] One or more detection channels may contain any suitable detector known in the art, such as a photomultiplier tube (PMT), charge-coupled device (CCD), and time-delay integration (TDI) camera. The detector may also contain non-imaging detectors or imaging detectors. If the detector is a non-imaging detector, each of the detectors may be configured to detect certain characteristics of light (e.g., intensity), but may not be configured to detect such characteristics as those varying depending on position within the imaging plane. Therefore, the output generated by each detector in each detection channel may be a signal or data, rather than an image signal or image data. In such examples, for example, the computer subsystem 36 of the inspection subsystem may be configured to generate an image of the sample from the non-imaging output of the detector. However, in other examples, the detector may be configured as an imaging detector, which is configured to generate an image signal or image data. Therefore, the inspection subsystem may be configured to generate images in several ways.

[0055] It should be noted that the information provided in this article... Figure 1 General descriptions may be included in the configuration of the inspection subsystem in the system embodiments described herein. Obviously, the configuration of the inspection subsystem described herein can be modified to optimize the performance of the inspection subsystem, as is typically done when designing a commercial inspection system. Alternatively, the system described herein can be implemented using an existing inspection subsystem (e.g., by adding the functionality described herein to an existing inspection system) (e.g., commercially available 29xx / 39xx series tools from KLA Corp., Milpitas, Calif). For some such systems, the methods described herein may be provided as optional functionality for the inspection system (e.g., in addition to other functions of the inspection system). Alternatively, the inspection subsystem described herein may be designed "from scratch" to provide a completely new inspection system.

[0056] Computer subsystem 36 may be coupled to the detector of the inspection subsystem in any suitable manner (e.g., via one or more transmission media that may include “wired” and / or “wireless” transmission media), such that the computer subsystem can receive the output generated by the detector. Computer subsystem 36 may be configured to perform several functions using the detector output. For example, the computer subsystem may be configured to detect events on a sample using the detector output. Detection of events on a sample may be performed by applying some defect detection algorithm and / or method (which may include any suitable algorithm and / or method known in the art) to the output generated by the detector. For example, the computer subsystem may compare the detector output with a threshold. Any output with a value higher than the threshold may be identified as an event (e.g., a potential defect), while any output with a value lower than the threshold may not be identified as an event.

[0057] The computer subsystem of the testing system may be further configured as described herein. For example, computer subsystem 36 may be part of, or configured to be part of, one or more computer subsystems described herein. Specifically, computer subsystem 36 may be configured to perform the steps described herein. Thus, the steps described herein may be performed "on the instrument" by a computer system or subsystem that is part of the testing system.

[0058] The computer subsystem of the testing system (and other computer subsystems described herein) may also be referred to herein as a computer system. Each of the computer subsystems or systems described herein may take various forms, including personal computer systems, graphics computers, mainframe computer systems, workstations, network devices, Internet devices, or other devices. Generally, the term "computer system" may be broadly defined to encompass any device having one or more processors that execute instructions from memory media. A computer subsystem or system may also include any suitable processor known in the art, such as a parallel processor. Additionally, a computer subsystem or system may include a computer platform with high-speed processing and software, as a standalone or networked tool.

[0059] If the system contains more than one computer subsystem, then the different computer subsystems can be coupled to each other, allowing images, data, information, instructions, etc., to be sent between the computer subsystems. For example, computer subsystem 36 can be coupled to, for example, via any suitable transmission medium. Figure 1 The computer subsystem 102, shown in dashed lines, may include any suitable wired and / or wireless transmission media known in the art. Two or more such computer subsystems may also be effectively coupled by sharing a computer-readable storage medium (not shown).

[0060] Although the above-described inspection subsystem is an optical or light-based inspection subsystem, in another embodiment, the inspection subsystem is configured as an electron beam inspection subsystem. In the electron beam inspection subsystem, the energy directed to the sample contains electrons, and the energy detected from the sample contains electrons. Figure 1a In the embodiment shown, the inspection subsystem includes an electron column 122, and the system includes a computer subsystem 124 coupled to the inspection subsystem. The computer subsystem 124 may be further configured as described above. Additionally, this inspection subsystem may be as described above... Figure 1 It is coupled to one or more computer subsystems in the same way as shown in the diagram.

[0061] For example Figure 1aAs shown, the electron column includes an electron beam source 126 configured to generate electrons focused onto a sample 128 by one or more elements 130. The electron beam source may include, for example, a cathode source or an emitter tip, and the one or more elements 130 may include, for example, a gun lens, an anode, a beam limiting aperture, a gate valve, a beam selection aperture, an objective lens, and a scanning subsystem, all of which may include any such suitable elements known in the art.

[0062] Electrons returning from the sample (e.g., secondary electrons) can be focused onto detector 134 by one or more elements 132. One or more elements 132 may include, for example, a scanning subsystem that may be included in the same scanning subsystem as element 130.

[0063] The electron column may include any other suitable element known in the art. Furthermore, the electron column may be further configured as described in U.S. Patent No. 8,664,594 to Jiang et al., issued April 4, 2014; U.S. Patent No. 8,692,204 to Kojima et al., issued April 8, 2014; U.S. Patent No. 8,698,093 to Gubbens et al., issued April 15, 2014; and U.S. Patent No. 8,716,662 to MacDonald et al., issued May 6, 2014, all of which are incorporated herein by reference as if fully set forth herein.

[0064] although Figure 1a The electron beam is shown as being configured such that electrons are guided to the sample at an oblique incident angle and scattered from the sample at another oblique incident angle, but the electron beam can be guided to and scattered from the sample at any suitable angle. Furthermore, the electron beam inspection subsystem can be configured to use multiple modes to produce outputs for the sample as further described herein (e.g., with different illumination angles, collection angles, etc.). These multiple modes of the electron beam inspection subsystem can differ in any output production parameters of the inspection subsystem.

[0065] Computer subsystem 124 may be coupled to detector 134 as described above. The detector detects electrons returning from the sample surface, thereby forming an electron beam image (or other output) of the sample. The electron beam image may include any suitable electron beam image. Computer subsystem 124 may be configured to use the output generated by detector 134 to detect events on the sample, said output may be performed as described above or in any other suitable manner. Computer subsystem 124 may be configured to perform any additional steps described herein. Figure 1a The inspection subsystem shown in the document can be further configured as described in this document.

[0066] It should be noted that the information provided in this article... Figure 1a General descriptions may be included in the configuration of the electron beam inspection subsystem in the embodiments described herein. As with the optical inspection subsystem described above, the configuration of the electron beam inspection subsystem described herein can be modified to optimize the performance of the inspection subsystem, as is typically done when designing commercial inspection systems. Alternatively, the system described herein can be implemented using existing inspection subsystems (e.g., by adding the functionality described herein to an existing inspection system) (e.g., commercially available tools from KLA). For some such systems, the methods described herein may be provided as optional functionality of the system (e.g., in addition to other system functions). Alternatively, the system described herein may be designed “from scratch” to provide a completely new inspection system.

[0067] Although the inspection subsystem described above is a light or electron beam inspection subsystem, it can also be an ion beam inspection subsystem. This inspection subsystem can be as follows: Figure 1a The configuration shown can be modified by replacing the electron beam source with any suitable ion beam source known in the art. Additionally, the inspection subsystem can include any other suitable ion beam imaging system, such as those found in commercially available focused ion beam (FIB) systems, helium ion microscopy (HIM) systems, and secondary ion mass spectrometry (SIMS) systems.

[0068] As further mentioned above, the testing subsystem can be configured to have multiple modes. Generally, a "mode" is defined by the parameter values ​​of the testing subsystem used to generate the output of the sample. Therefore, different modes may differ in the values ​​of at least one of the optical or electron beam parameters of the testing subsystem (except for the position on the sample where the output or image is generated). For example, for a light-based testing subsystem, different modes may use light of different wavelengths. For different modes, the wavelengths of the light directed to the sample, as further described herein (e.g., by using different light sources, different spectral filters, etc.), may differ. In another embodiment, different modes may use different illumination channels. For example, as mentioned above, the testing subsystem may include more than one illumination channel. Thus, different illumination channels can be used for different modes.

[0069] Multiple modes may also differ in illumination and / or collection / detection. For example, as further described above, the inspection subsystem may include multiple detectors. Thus, one detector may be used in one mode, and another detector may be used in another mode. Furthermore, modes may differ from each other in more than one way as described herein (e.g., different modes may have one or more different illumination parameters and one or more different detection parameters). The inspection subsystem may be configured to scan samples using different modes in the same scan or different scans, depending, for example, on the ability to scan samples simultaneously using multiple modes.

[0070] The system described herein can be configured as another type of semiconductor-related process / quality control system, such as a defect inspection system and a measurement system. For example, the systems described and shown herein... Figure 1 and 1a Implementations of the system may be modified in one or more parameters to provide different imaging capabilities depending on the application in which it will be used. In one embodiment, the electron beam inspection subsystem configuration described herein may be modified to configure an electron beam defect inspection system. For example, Figure 1a The subsystems shown can be configured to have higher resolution if they are used for defect inspection or measurement rather than for testing. In other words, Figure 1 and 1a The embodiments of the systems shown describe some general and various configurations of the subsystems, which can be customized in several ways as will be known to those skilled in the art to produce subsystems with different imaging capabilities that are more or less suitable for different applications.

[0071] As mentioned above, the examination subsystem can be configured to direct energy (e.g., light, electrons) to a physical version of the sample and / or scan the energy on the physical version of the sample, thereby producing an actual image of the physical version of the sample. In this way, the examination subsystem can be configured as an "actual" imaging system rather than a "virtual" system. Figure 1 The storage media (not shown) and computer subsystem 102 shown herein can be configured as a “virtual” system. Systems and methods configured as “virtual” verification systems are described in U.S. Patent No. 8,126,255, jointly assigned by Bhaskar et al., published February 28, 2012, and U.S. Patent No. 9,222,895, jointly assigned by Duffy et al., published December 29, 2015, both of which are incorporated herein by reference as if fully set forth herein. The embodiments described herein can be further configured as described in those patents.

[0072] One or more components executed by one or more computer subsystems contain Figure 1 The trained machine learning (ML) model 106 shown in the figure is configured to transform images of samples generated by the testing subsystem into higher resolution images of the samples.

[0073] Figure 2 This section typically provides examples of design data, test sample images, and higher-resolution images to illustrate how a trained ML model can be useful. Specifically, Figure 2Design image 200 shows a portion of the sample's design data and the locations of locatable DOIs (e.g., tip-to-tip bridges) on it (within circles superimposed on the design image). Design image 202 shows the same portion of the design data and the locations of locatable perturbation points, such as CD variations (within circles superimposed on design image 202).

[0074] Image 204 illustrates a possible image that the testing subsystem can generate for a sample, where a tip-to-tip bridge type DOI can be detected at the location circled on the superimposed image. Image 206 illustrates a possible image that the testing subsystem can generate for a sample, where a CD variation type of scrambling point can be detected at the location circled on the superimposed image. As can be seen by comparing the circled portions of images 204 and 206, the DOI and scrambling point have substantially similar signal characteristics in both images, making it difficult to separate the DOI from the scrambling point. Specifically, when the DOI and scrambling point have substantially similar signal characteristics, it is practically impossible to detect the DOI without detecting the scrambling point, and it is impossible to determine which of the detected events is the DOI and which is the scrambling point without additional information.

[0075] However, the two images can be input into a trained ML model as described herein, thereby transforming them into higher resolution images. For example, the trained ML model can transform image 204 into a higher resolution image 208 of the sample and image 206 into a higher resolution image 210 of the sample. As can be seen by comparing the circled portions of image 208 and image 210 corresponding to images 204 and 206 respectively, the DOI in image 208 has a signal that is significantly different from the scrambling point signal in image 210. In this way, the signals of DOI and scrambling points in the higher resolution image generated by the trained ML model can have sufficient distance so that the DOI can remain detected while filtering out scrambling points. As further described herein, the inventors have found that using higher resolution images generated from images produced by the inspection subsystem by a trained ML model can reduce the number of scrambling points in the inspection results by up to 50 times, which has significant benefits for inspection and other related processes (such as defect inspection).

[0076] Each of the design images, test images, reference images, lower-resolution images, actual higher-resolution images, simulated higher-resolution images, etc., shown in the accompanying drawings is not intended to illustrate any particular sample or characteristic thereof that can be used with the embodiments described herein. Similarly, each of the lower and higher-resolution images shown in the accompanying drawings is not intended to illustrate any particular actual or simulated image that can be generated for a sample. Rather, the design images, test images, reference images, lower-resolution images, actual higher-resolution images, simulated higher-resolution images, etc., shown in the accompanying drawings are intended only to facilitate an understanding of the embodiments described herein. The images from the actual input and output of the trained ML model will vary depending on the sample and its characteristics (related to its design) and the configuration of the imaging system that generates the actual image of the sample (used to train the ML model, thereby affecting the simulated image generated by the trained ML model).

[0077] The embodiments described herein are configured to perform two training phases. The first phase is pre-training using only simulated data from the design. For example, one or more computer subsystems are configured to pre-train an initial ML model using a pre-training set, thereby generating a pre-trained ML model. The pre-training set contains simulated test images of test samples designated as pre-training inputs and corresponding simulated higher-resolution images of test samples designated as pre-training outputs. In the pre-training phase, the simulated test images may include simulated test images (which may have simulated defects and / or other variations such as LER) and corresponding simulated reference images (e.g., images of defect-free samples). The design of the samples may also be used as pre-training inputs in the pre-training phase. The simulated higher-resolution images used in the pre-training phase may be higher-resolution images that may contain simulated defects and other variations such as LER. In the pre-training phase, the initial ML model can learn to transform from the pre-training inputs to the pre-training outputs using the pre-training inputs and outputs.

[0078] Figure 3This embodiment demonstrates a pre-trained approach for improving resolution networks. In this embodiment, simulated test images 300, simulated reference images 302, and designs 304 for test samples are used as pre-training inputs to an initial ML model 306. During the pre-training phase, simulated higher-resolution images 310 are used as pre-training outputs. In this way, when the pre-training inputs are fed into the initial ML model, the ML model outputs a higher-resolution image 308. In step 312, one or more computer subsystems compare the pre-training output (i.e., the simulated higher-resolution image 310) with the simulated output (i.e., the high-resolution image 308). As shown in step 314, one or more computer subsystems may determine, based on the comparison, whether one or more initial model parameters should be adjusted. If it is determined that one or more initial model parameters should be adjusted, then one or more computer subsystems may determine the adjusted initial model parameters, as shown in step 316, which may be determined based on any differences between the pre-training output and the simulated output. The determined one or more initial model parameters may then be applied to the initial ML model. If it is determined that no adjustment is needed for one or more initial model parameters, then one or more computer subsystems may designate the current version of the initial ML model as a pre-trained ML model, as shown in step 318.

[0079] In this way, Figure 3 The steps illustrated herein can be performed in a feedback loop until the initial ML model produces an output that substantially matches the pre-training output. In other words, pre-training may involve feeding pre-training inputs into the initial ML model and changing one or more parameters of the initial ML model until the output produced by the initial ML model matches (or substantially matches) the pre-training output. Pre-training may involve changing any one or more trainable parameters of the initial ML model. For example, one or more parameters of the initial ML model trained by the embodiments described herein may include one or more weights for any layer of the initial ML model with trainable weights. In this example, the weights may include weights for pooling layers but not weights for pooling layers. Figure 3 The steps shown can be performed offline during the setup phase of the test.

[0080] Therefore, the pre-training phase can be performed using only simulated pre-training inputs and outputs, a novel feature of the embodiments described herein. In other words, none of the images used in the pre-training phase are actual (i.e., non-simulated) images of the samples generated using physical versions of the samples themselves and some imaging hardware. Therefore, an advantage of the embodiments described herein is that using simulated optical (or other) images to pre-train the network reduces the tooling time required to acquire optical or other lower-resolution inspection images and high-resolution images (e.g., scanning electron microscope (SEM) images).

[0081] In one embodiment, the initial ML model is a generative network. A “generative” network can generally be defined as a model with probabilistic properties. In other words, a “generative” network is not a network that performs forward simulation or rule-based methods; therefore, a physical model of the processes involved in generating the actual images is unnecessary. Instead, as further described herein, the generative network can be learned based on a suitable training set of data (because its parameters can be learned). The generative network can be configured with a deep learning (DL) architecture that can contain multiple layers performing several algorithms or transformations. The number of layers included in the generative network can be left to the use case. For practical purposes, a suitable range of layers is from 2 to dozens. A deep generative network that learns the joint probability distribution (mean and variance) between lower-resolution, inspection-type sample images (e.g., images of actual wafers) and higher-resolution sample images (e.g., SEM or defect inspection-type images) can be configured as further described herein.

[0082] In another embodiment, the initial ML model is a Generative Adversarial Network (GAN). Generally, a GAN consists of two adversarial models: a generative model G for capturing the data distribution and a discriminative model D for estimating the probability that a given sample comes from the training data rather than G. G and D can be multilayer perceptrons, i.e., nonlinear mapping functions. The generator builds from the prior noise distribution P. z (z) to data space G(z; θ) g The mapping function is used to learn the generator distribution P on the data x. g Where G is a function with parameter θ g The multilayer perceptron represents a differentiable function. The generator is trained to produce images indistinguishable from real images. An adversarially trained discriminator is trained to detect forgeries created by the generator. Both the generator and discriminator are trained as well as possible so that the generator produces extremely well-made "forged" images. For example, the generator outputs synthetic samples given a random noise variable input z. Over time, the generator is trained to capture the real data distribution by causing the discriminator to reject images it considers bad forgeries.

[0083] Additional descriptions of the general architecture and configuration of GANs can be found in: Goodfellow et al., “Generative Adversarial Nets,” arXiv:1406.2661, June 10, 2014, p. 9; Kingma et al., “Semi-supervised Learning with Deep Generative Models,” NIPS 2014, October 31, 2014, pp. 1–9; Mirza et al., “Conditional Generative Adversarial Nets,” arXiv:1411.1784, November 6, 2014, p. 7; and Makhzani et al., “Adversarial Autoencoders.” "Autoencoders", arXiv:1511.05644v2, May 25, 2016, p. 16; and Isola et al., "Image-to-Image Translation with Conditional Adversarial Networks", arXiv:1611.07004v2, November 22, 2017, p. 17; which are incorporated herein by reference as if fully set forth herein. The embodiments described herein may be further configured as described in these references.

[0084] In another embodiment, the initial ML model is an autoencoder. An autoencoder, autoassociator, or Diablo network is an artificial neural network used for unsupervised learning of efficient encoding. The purpose of an autoencoder is to learn a representation (encoding) of a set of data, often for dimensionality reduction. Recently, the concept of autoencoders has been increasingly used to learn generative models of data. Architecturally, the simplest form of an autoencoder is a feedforward, non-recursive neural network very similar to a multilayer perceptron (MLP), having an input layer, an output layer, and one or more hidden layers connecting them, but where the output layer has the same number of nodes as the input layer and aims to reconstruct its own input (rather than predicting a target value for a given input). Therefore, an autoencoder is an unsupervised learning model. An autoencoder always consists of two parts: an encoder and a decoder. Various techniques exist to prevent autoencoders from learning identity functions and to improve their ability to capture important information and learn richer representations. Autoencoders can include any suitable variant of an autoencoder, such as noise-reducing autoencoders, sparse autoencoders, variational autoencoders, and compressed autoencoders.

[0085] Variational autoencoders (VAEs) are components that leverage the advantages of deep learning (DL) and variational inference, leading to significant advancements in generative modeling. Alternatively or concurrently, VAEs combined with GANs or deep generative adversarial networks (DGANs) can be configured as described in Mahzani et al., "Adversarial Autoencoders," arXiv:1511.05644v2, May 25, 2016, page 16 (which is incorporated herein by reference as if fully set forth herein). The embodiments described herein can be further configured as described with reference to the embodiments described herein.

[0086] In one embodiment, one or more computer subsystems are configured to generate simulated inspection images and simulated higher-resolution images from the design of the test sample. For example, the computer subsystem may be configured to select design clips at random locations across the mask. The simulated data can be generated in pairs of two resolutions. The first matches the data acquired by the inspection tool. The second is the target improved resolution. Many such pairs can be simulated to pre-train a generative network without the need for expensive pairs of actual sample images (e.g., optical and SEM data), including in some cases where SEM is the only available source for higher-resolution imaging.

[0087] In this embodiment, generating simulated test images and simulated higher-resolution images includes simulating process variations in the design formed on the test sample by applying a perturbation representing the expected process variation to the design to generate a perturbed design. In this way, the computer subsystem can be configured to simulate process variations as design perturbations. Therefore, a novel feature of the embodiments described herein is that pre-training can be performed using image simulations from the perturbed design. Process variation simulations can be performed to simulate variations in one or more characteristics of the design printed on the sample, such as LER variations from the design, CD variations from the design, etc.

[0088] Process variations can simulate only the nominal perturbations that would occur to the design formed on the sample. In this way, the perturbed design can represent what the design on the sample would look like if it were formed using the optimal or nominal process parameters (i.e., if the process were to operate normally). However, perturbed designs can also, or alternatively, generate for non-nominal process variations to simulate how the design is formed on the sample using other or non-nominal process parameters, which may be within or even outside the known or expected process window used to form the design on the sample. In this way, perturbations can be applied to the design to generate an image representing how the design is formed on the sample under one or more process parameters, thereby generating one or more perturbed designs. Thus, applying perturbations in this way can simulate an image that can be generated by a process window defined (PWQ) type process or by a wafer forming and imaging focused exposure matrix (FEM) type. Examples of the PWQ method are described in U.S. Patent No. 6,902,855 to Peterson et al., issued June 7, 2005; U.S. Patent No. 7,418,124 to Peterson et al., issued August 26, 2008; U.S. Patent No. 7,729,529 to Wu et al., issued June 1, 2010; U.S. Patent No. 7,769,225 to Kekare et al., issued August 3, 2010; U.S. Patent No. 8,041,106 to Pak et al., issued October 18, 2011; U.S. Patent No. 8,111,900 to Wu et al., issued February 7, 2012; and U.S. Patent No. 8,213,704 to Peterson et al., issued July 3, 2012, all of which are incorporated herein by reference as if fully set forth herein. The embodiments described herein may include any steps of any method described in these patents and may be further configured as described in these patents, except that they differ from some of the methods and systems described in these patents. The embodiments described herein can simulate PWQ or FEM wafers by applying different perturbations to the design rather than printing the design on a physical wafer with different process parameters.

[0089] Figure 4This describes an embodiment generated using simulated offline pre-trained data. As shown in step 400, one or more computer subsystems may select a design clip 400 as described above. The selection of the design clip can be performed in any suitable manner from any of the designs, design data, or design information described herein. The design clip can have any suitable size and can be selected in any suitable manner. Although for clarity, Figure 4 Only design clips are shown herein, but the embodiments described herein will generally select many more design clips, and the number of design clips selected may be determined based on the amount of simulated pre-training data expected to be required during the pre-training phase. The patterns shown in design clip 400 are not intended to illustrate any particular design or sample that can be used with the embodiments described herein. In other words, the embodiments described herein are not specific to any design or sample.

[0090] As shown in step 402, one or more computer subsystems can introduce process variations into the design clip. In some such embodiments, a Gaussian process model is used to perform the application of perturbations to the design. For example, the perturbation may represent an actual variation and can be implemented using a Gaussian process model. A Gaussian process model may include any such suitable model known in the art. Additionally, while a Gaussian process model may be particularly well-suited for applying perturbations in the embodiments described herein, any other suitable process model may be used for this step (e.g., when a different process model can better estimate the perturbations occurring on the design when formed on a sample). The perturbation design can be suitable for simulating higher resolution images. For example, as... Figure 4 As shown, the output of the process change step 402 can be a simulated SEM image 404.

[0091] In another embodiment, generating the simulated verification image further includes transforming the perturbation design into the image domain. For example, the perturbation design may produce a binary version of the original design, as described above, which can be suitable for simulating a higher resolution image (e.g., simulating SEM image 404). Therefore, this binary version of the original design still cannot represent the image to be generated by the imaging tool. In some examples, the binary version of the perturbation design may be transformed into a grayscale image that more accurately represents the image to be generated by the imaging tool. In this way, the perturbation design can be transformed from the design domain into the image domain of the higher resolution imaging tool.

[0092] Similarly, the perturbation design can be transformed into the image domain of a lower-resolution imaging tool. As further described herein, the transformation into the image domain of a lower-resolution imaging tool is not as simple as transforming a binary image into a grayscale image, because the lower-resolution imaging tool can produce images that look different from the design and the perturbation design (e.g., from...). Figure 4The differences between the design clip 400, the simulated SEM image 404, the simulated test image 408, and the simulated reference image 412 shown in the figure can be seen.

[0093] In another embodiment, a partially coherent model is used to transform the perturbation design into the image domain. For example, a partially coherent model (PCM) can be used to transform the perturbation design into the optical domain. In this example, such as Figure 4 As shown, for example, the perturbation design of the simulated SEM image 404 can be input to use PCM reproduction performed in step 406. Reproduction using the PCM step can produce a simulated test image 408 that can serve as a simulated verification test image. One or more computer subsystems can also be configured to average the reproduced image from PCM as shown in step 410 to produce a simulated reference image 412. Averaging the reproduced image from PCM can be performed in any suitable manner known in the art and can be performed using multiple simulated test images to better simulate the possible appearance of the defect-free reference image. Then, pre-training can be performed offline using the design and results of the PCM and LER simulations during the setup phase, as further described herein.

[0094] PCMs can have any suitable configuration and parameters known in the art. Generally, PCMs are particularly well-suited for approximate optical systems, such as those described herein. However, if a different model is more suitable for approximating (simulating) an imaging tool used to generate a test image, then that model can be used to transform the perturbation design into the image domain. For example, a PCM model may be suitable for some, but not all, optical imaging tools. Additionally, a PCM model may not be suitable for approximate electron beam imaging tools. However, depending on the configuration of the imaging tool, the user can select, configure, or adapt an appropriate model to transform the perturbation design into the image domain.

[0095] In some cases, the perturbation design (whether in its binary form or transformed into a grayscale image) may not be suitable as a simulated high-resolution image for test samples. In such situations, a PCM model or a different model can be applied to the perturbation design to transform it into the image domain of a higher-resolution imaging tool. If a simulated high-resolution image is generated in this way, the models used to transform the perturbation design into a simulated high-resolution image and to simulate a lower-resolution image can be of the same type but with different parameters, such that different models approximate different imaging tools or different imaging resolutions of the same tool. However, different types of models can be used to transform the perturbation design into different image domains, whether it is a different type of image at a different resolution or a different type of image at a different resolution. For example, a PCM model can be used to transform the perturbation design into a simulated lower-resolution image, and a non-PCM model can be used to transform the perturbation design into a simulated higher-resolution image.

[0096] The second training phase, performed by the embodiments described herein, can be retraining using a limited set of real or non-simulated sample images (e.g., relatively limited optical and SEM tool data). For example, one or more computer subsystems are configured to retrain a pre-trained ML model using a training set, thereby generating a trained ML model. The training set includes images generated by a validation subsystem for test samples, designated as training inputs, and corresponding higher-resolution images of the test samples generated by a high-resolution imaging system, designated as training outputs. In this way, the embodiments described herein can improve the model performance of the validation layer by retraining using acquired low-resolution sample images, pairs of higher-resolution sample images (e.g., SEM images), and design data. The SEM tool data can be higher-resolution real-world data.

[0097] In the retraining phase (which may also be simply referred to as the training phase in the art), the training input may include test and reference images generated by the verification subsystem for the test samples. In some embodiments, the training set includes the design of the samples. For example, the training input may also include the design of the test samples, depending on the situation. The training output may include, for example, higher resolution images generated by SEM. During the training phase, the pre-trained ML model learns to transform from lower resolution training input to higher resolution training output. In this way, the training phase can utilize a limited amount of real-world tool data and inference. As further described herein, the trained ML model can then be used during verification to predict higher resolution images from the verification images.

[0098] In one embodiment, the number of images in the training set is less than the number of simulated test images in the pre-training set, and the number of higher-resolution images in the training set is less than the number of simulated higher-resolution images in the pre-training set. For example, an additional advantage of the embodiments described herein is that, as described herein, the pre-trained network minimizes the number of pairs of training inputs and outputs (e.g., optical and high-resolution SEM images) required for training.

[0099] Figure 5An embodiment using limited tool data for retraining is illustrated. In this embodiment, the actual test image 500, the actual test reference image 502, and the design 504 for the test samples are used as training inputs to the pre-trained ML model 506. During the training phase, a higher resolution image 510 of the samples is used as the training output. In this way, when the training input is fed into the pre-trained ML model, the pre-trained ML model will output a higher resolution image 508. In step 512, one or more computer subsystems compare the training output (i.e., the higher resolution image 510 of the test samples) with the simulated output (i.e., the higher resolution image 508). As shown in step 514, one or more computer subsystems may determine whether one or more pre-trained ML model parameters should be adjusted based on the comparison. If it is determined that one or more pre-trained ML model parameters should be adjusted, then one or more computer subsystems may determine the adjusted pre-trained ML model parameters, as shown in step 516, which may be determined based on any differences between the training output and the simulated output. Next, the determined pre-trained ML model parameters can be applied to the pre-trained ML model. If it is determined that no adjustment of the pre-trained ML model parameters is required, then one or more computer subsystems can designate the current version of the pre-trained ML model as the trained ML model, as shown in step 518.

[0100] In this way, Figure 5 The steps illustrated herein can be performed in a feedback loop until the pre-trained ML model produces an output that substantially matches the training output. In other words, training can involve feeding training inputs into the pre-trained ML model and changing one or more parameters of the pre-trained ML model until the output produced by the pre-trained ML model matches (or substantially matches) the training output. Training can involve changing any one or more trainable parameters of the pre-trained ML model. For example, one or more parameters of the pre-trained ML model trained by the embodiments described herein can include one or more weights for any layer of the pre-trained ML model with trainable weights. In this example, the weights may include weights for pooling layers but not weights for pooling layers. Figure 5 The steps shown can be performed offline during the setup phase of the test.

[0101] In another embodiment, the simulated inspection image, the simulated higher-resolution image, the image generated for the test sample, and the higher-resolution image of the test sample all comprise only optical images. For example, all images described herein may be optical images. When the inspection subsystem is configured as a light-based tool (e.g., as...), Figure 1This embodiment may be suitable when simulating a higher resolution image means simulating an image that will be produced by a higher resolution optical tool (whether it is an inspection subsystem operating in a higher resolution mode or another tool, such as a measurement or defect inspection tool configured to be based on light imaging).

[0102] In another embodiment, the simulated inspection image and the image generated for the test sample comprise only optical images, and the simulated higher-resolution image and the higher-resolution image of the test sample comprise only electron beam images. In this way, some of the images described herein can be optical images, and others can be electron beam images. This is the most common use of the embodiments described herein. For example, the embodiments described herein are particularly suitable for light-based inspection subsystems (e.g., Figure 1 Used together (as shown in the diagram), the light-based inspection subsystem can be advantageously used to scan samples at relatively low resolution with generally high throughput. For such inspections, events or defects detected by the inspection are typically imaged using a higher-resolution electron beam tool (e.g., a SEM defect inspection tool), which can be configured as follows: Figure 1a (As shown in the illustration). Therefore, a popular way to configure and use the embodiments described herein is to train and use an ML model with a lower resolution optical inspection image as input and a higher resolution electron beam defect inspection type image as output. In this way, the ML model can be trained and then used to generate a higher resolution SEM type image from a lower resolution optical inspection type image.

[0103] In another embodiment, the simulated inspection image, the simulated higher-resolution image, the image generated for the test sample, and the higher-resolution image of the test sample all contain only electron beam images. For example, all images described herein may be electron beam images. When the inspection subsystem is configured as an electron beam-based tool (e.g., such as...), Figure 1a This embodiment may be suitable when simulating a higher resolution image means simulating an image that will be produced by a higher resolution electron beam tool (whether it is an inspection subsystem operating in a higher resolution mode or another tool, such as a measurement or defect inspection tool configured to be based on electron beam imaging).

[0104] Although the embodiments described herein are described with respect to a trained ML model, the embodiments are not limited to a single trained ML model. For example, different ML models can be trained as described herein to produce images or other outputs for different modes. Specifically, in most cases, different modes of an inspection or other imaging tool will produce images and / or outputs that differ from each other in one of several possible ways (e.g., noise level, contrast, resolution, image type (e.g., DF vs. BF, optical vs. electron beam, etc.) and the like). Therefore, if an ML model is trained to generate a simulated higher-resolution image of a tool's high-resolution mode from a low-resolution mode of the tool or a lower-resolution image generated by a different tool, it is likely that it will not be suitable for generating a higher-resolution image from another high-resolution mode of the tool or from another low-resolution mode of the tool or a lower-resolution image generated by a different tool. Thus, multiple ML models can be trained individually and independently, one for each mode and combination of interest. For example, a first ML model can be trained as described herein to generate a higher-resolution SEM image from a pattern of SEMs of lower-resolution images generated by a first pattern from the verification tool, and a second ML model can be trained as described herein to generate a higher-resolution SEM image from the same or different patterns of SEMs of lower-resolution images generated by a second pattern from the verification tool. However, the same initial ML model can be used for each pattern, although this is not necessary. Each trained ML model can then be used to generate pattern-specific datasets.

[0105] The test samples used to generate the pre-training set and training set may differ from the test samples of the higher-resolution images generated by the trained ML model. Furthermore, although this document describes the pre-training set and training set in relation to test samples, more than one test sample can be used to generate both. However, generally, since the pre-training set only contains simulated images, no actual samples are required to generate it. In this way, one or more test samples can be used to generate the pre-training set and retraining set, and higher-resolution simulated images can generate "runtime" samples. The test samples and runtime samples can be of the same type, for example, they may have the same design and can be processed using the same process steps, although this is not always the case, although it is further described herein. Additionally, the trained ML model can be used to generate more than one runtime sample of simulated images, all of which may have the same design and can be processed using the same process steps.

[0106] Generally, one or more test samples used to generate pre-training and retraining sets are used to pre-train and retrain the ML model, respectively. Samples used by the trained ML model to simulate higher-resolution images can have the same design and be processed in the same process (and therefore have the same "layers"). Training the ML model in this way ensures that the simulated higher-resolution images will most closely approximate the actual higher-resolution images used for the samples. However, in some cases, the test samples may have sufficiently similar characteristics to the running-time samples (e.g., patterned features, materials, etc.), so that even if the test samples and running-time samples do not have the same design, the ML model trained on the test samples can still be used to generate higher-resolution simulated images of the running-time samples. In such cases, the higher-resolution simulated images generated by the trained ML model should be identical to the trained imaging modality. In other words, as further described herein, an ML model trained to generate higher-resolution simulated images from an imaging modality may not be suitable for generating higher-resolution simulated images from another imaging modality. Therefore, if two samples that have at least some similarity in at least a part of their design are imaged in the same way or will be imaged in the same way, then one of the samples can be used as a test sample, and a trained ML model can be used to generate a higher resolution simulated image of the other sample.

[0107] In this way, a trained ML model can be reused to generate higher-resolution simulated images of samples for which it did not need to be trained. In this example, if two different samples with two different designs share at least some patterned features in a portion of a design (e.g., a region of a similar memory array) formed of similar materials and of the same or similar size, then an ML model trained for one of the samples can generate a higher-resolution simulated image of a portion of the design of the other sample. Even if the ML model trained for one sample cannot generate a higher-resolution simulated image for another sample with a different design, if there is some similarity between the samples, then the trained ML model can be used as a starting configuration for retraining another sample to build a different trained ML model. This retraining can be performed as described herein.

[0108] One or more computer subsystems are configured to input images generated for samples by an inspection subsystem during inspection into a trained ML model, thereby generating higher-resolution images of the samples. In this way, the trained ML model is used to improve the resolution of inputs (e.g., optical images) at candidate defect locations. Therefore, another advantage of the embodiments described herein is that, since no additional sample scanning is required, the embodiments can generate higher-resolution images without affecting output. Additionally, the embodiments described herein enable inspection to run at a larger pixel size, but with the advantage of a smaller pixel size (because higher-resolution images can be generated from inspection images produced at a lower resolution (i.e., a larger pixel size) using the trained ML model described herein). Enabling lower-resolution inspection while retaining the advantages of higher-resolution inspection can also result in a significant increase in output (e.g., output increased by more than 3 times). The trained ML-based model can also be used to generate higher-resolution images for samples in real time during on-tool inspection. In this way, higher-resolution images can be generated for samples from online inspection images during runtime.

[0109] The computer subsystem can be configured to input images generated for samples into a trained ML model in any suitable manner known in the art. Although some embodiments herein are described (for clarity and simplicity only) as generating “higher resolution images” for samples, the trained ML model described herein can be used to generate any number of higher resolution images of a sample limited only by the images of the samples input to the trained ML model. Furthermore, although some embodiments herein are described (for clarity and simplicity) as generating higher resolution images for “samples”, the embodiments described herein are not limited to generating higher resolution images only for samples.

[0110] The images input from the computer subsystem to the trained ML model can include test and reference images at any location on the sample. The test and reference images can include, for example, images generated at corresponding locations in different grains of the sample, different cells on the sample, different mask fields on the sample, etc. The test images can also include images of the sample generated by the verification subsystem, and the reference images can include corresponding portions of a standard reference image stored in a computer-readable medium. The standard reference image may or may not be generated by imaging the actual sample.

[0111] While inputting test and reference images of locations on the sample into a trained ML model can improve the quality of the higher-resolution images generated by the trained ML model for those locations, this is not necessary. For example, any sample images generated by the inspection subsystem during inspection can be input one by one into the trained ML model, and the higher-resolution image generated for each location can then be used for defect detection. In this way, the inspection process performed on the sample may not include defect detection performed on the sample images generated from scanning the sample during inspection, but may include defect detection performed using higher-resolution images generated from lower-resolution images produced by scanning. Whether this is a practical solution for inspection may depend on the operations required to generate the higher-resolution images and the speed at which the higher-resolution images are generated. Defect detection performed using high-resolution images may include standard grain-to-grain defect detection or grain-to-database detection, where the reference image is a stored higher-resolution reference image. In other words, simulated higher-resolution images can be input into any suitable defect detection method or algorithm, as currently performed in the art. However, such high-resolution images can also be suitable for single-image detection, as described in U.S. Patent No. 10,186,026 to Karsenti et al., issued January 22, 2019, which is incorporated herein by reference as if fully set forth herein.

[0112] In one embodiment, one or more computer subsystems are configured to input the design of a sample along with images generated for the sample by an inspection subsystem during inspection into a trained ML model. For example, the inspection subsystem may generate a lower-resolution image of the sample at the system's Nyquist rate. In such cases, new information cannot be obtained solely from the lower-resolution image. Any super-resolution technique attempting to reconstruct information lost during imaging can produce artifacts. The idea behind adding design and simulating process variations is to add information to the system, and the ML model learns what the actual perturbations are and the process variations that cause them. Therefore, a novel feature of the embodiments described herein is that they can use the design as an additional channel to aid in the reconstruction of higher-resolution images. Furthermore, the advantages of the embodiments described herein are that they can utilize alignment design information readily available on many currently used optical inspection tools, adding an additional information channel to the high-resolution reconstruction method.

[0113] Figure 6An embodiment is shown in which images generated by a testing subsystem for a sample during testing are input into a trained ML model to generate a higher-resolution image of the sample. In this embodiment, one or more computer subsystems input the actual test image 600 of the sample, the actual reference image 602 of the sample, and the design 604 of the sample into a trained ML model 606. The trained ML model transforms the input into a reconstructed higher-resolution image 608, which can then be used for one or more additional functions described herein.

[0114] One or more computer subsystems are configured to determine information about the sample from the resulting higher-resolution image. The information about the sample may include determining the events detected on the sample during inspection, such as whether the detected event is a decoy or an actual defect (as described herein with decoy filtering) and the classification of detected events or defects not filtered by decoy filtering (as described herein with defect classification).

[0115] The information used to determine a sample may also, or alternatively, include the ability to perform similar measurements on the sample and / or on events or defects detected on the sample. For example, simulated high-resolution images generated by a trained ML model can be input into a measurement method or algorithm, in which various properties of detecting events, defects, or patterned features (defect or not) in the simulated high-resolution image are determined or measured. Such features may include any properties that can be determined from the high-resolution image, such as CD, LER, shape properties, spatial relationships between defects and patterned features, or spatial relationships between patterned features, etc.

[0116] Higher resolution images are more useful for alignment-to-design applications than images generated by the inspection subsystem. For example, higher resolution images can be aligned with the design of a sample using higher accuracy and precision by alignment-to-design methods or algorithms, such as the Pixel-to-Design Alignment (PDA) algorithm used by some commercially available tools from KLA. Therefore, the information determined for a sample using a higher resolution image can include defect coordinates in the design data space, rather than the sample space or tool-specific coordinates reported by the inspection subsystem. Information about the location of defects relative to the design, determined from simulated higher resolution images, can also be used to determine additional information about the sample, such as information about weaknesses or hotspots in the design (i.e., locations where defects appear or tend to repeat in multiple instances of the same part of the design).

[0117] Therefore, in general, simulated high-resolution images generated by the trained ML model described herein can be used as input to any method or algorithm that can be used to determine information about any sample described herein from a relatively high-resolution image of the sample.

[0118] In one embodiment, the information for determining a sample from the generated higher-resolution image includes performing obfuscation point filtering of defects detected by inspection based on the generated higher-resolution image, thereby generating a filtered defect population for the sample. In this way, the improved higher-resolution image can be used for obfuscation point filtering. Therefore, the embodiments described herein help address the problem of future optical inspections becoming limited to obfuscation points. This means that previously there was sufficient separation between the signals of the DOI and the obfuscation points, so a threshold could be used to detect only the DOI without detecting obfuscation points. However, in many current and future scenarios, any threshold used to detect the DOI will necessarily detect at least some obfuscation points. In other words, the signal corresponding to the DOI (or any signal used for defect detection) cannot be sufficiently separated from the signal of the obfuscation points. Furthermore, in many cases, not only some obfuscation points are detected along with the DOI. Instead, extremely high obfuscation point rates are becoming increasingly common, where several obfuscation points are detected for each DOI, or even more obfuscation points are detected than the number of DOIs. Detecting such high-level obfuscation points renders the resulting data essentially useless.

[0119] Therefore, converting high-disruption-rate test results into actionable data may require reducing the disruption rate by up to 50 times the current disruption rate. For example, if a care area-based test produces 10... 7 If the current scrambling point event filtering reduces the number of events by 20 times, then there are still up to 500K total events available for sampling / defect inspection. In contrast, if the embodiments described herein reduce the total number of events after detection by 1000 times (meaning a 50-fold improvement over the current scrambling point filtering), then the total number of events available for sampling / inspection would be only 10K, significantly lower than 500K. The embodiments described herein provide such advanced scrambling point rates by using simulated higher-resolution images and utilizing design and optical simulations. Scrambling point filtering can be performed in any suitable manner known in the art. For example, the embodiments described herein provide the aforementioned improved scrambling point filtering simply by using the currently used scrambling point filtering with simulated higher-resolution images generated by the trained ML model described herein instead of lower-resolution inspection images (or by applying the currently used scrambling point filtering to simulated higher-resolution images).

[0120] As design rules continue to shrink, optical inspection becomes a finite number of obfuscation points. This limitation leads front-end semiconductor manufacturers to increasingly rely on electron microscopy for inspection, which comes with its own set of drawbacks. To reduce obfuscation point detection in optical inspectors to an acceptable level, the obfuscation point rate may need to be reduced by as much as 50 times or more compared to current rates. The embodiments described herein achieve a significant reduction in obfuscation points while maintaining, and even achieving, high sensitivity in optical inspection tools.

[0121] Figure 7 This describes an embodiment of on-tool inspection that can be performed by the embodiments described herein. In this embodiment, the sample is a wafer 700, which is loaded into an inspection tool 702 that performs on-tool inspection of the sample, wherein the sample is scanned by the tool 702 and the images generated by the scan are used to detect defects on the sample. The output generated by the defect detection step may include an initial batch result 704, which may contain a large number of detection events (e.g., 1 million to 2 million detection events). The detection events may include DOIs and scrambling points (and often even more scrambling points than DOIs), which may be separated from each other in additional steps.

[0122] The embodiments described herein may input information from initial batch results into a trained ML model 706, which may be configured as described herein. Specifically, test images generated for events (which may have corresponding reference images and designs) may be input into the trained ML model that generates higher-resolution images 708 of detected events. One or more computer subsystems may or may not input test images (and other optional inputs) of all detected events on the sample. For example, detected events in the initial batch results may be sampled in some way, and then test images of the sample set used only for detecting events may be input into the trained ML model.

[0123] In any case, the higher-resolution image 708 generated by the trained ML model can be input into a scrambling point filtering step 710, which can be performed as described herein. The output 714 of the scrambling point filtering step may contain any detection events that were not filtered out as scrambling points. As schematically shown by a wafer diagram representing the output of the scrambling point filtering step, the number of detection events remaining after scrambling point filtering may be substantially less than the number of detection events reported by the tool.

[0124] In another embodiment, the information used to determine the sample from the generated higher-resolution image includes classifying defects detected by inspection based on the generated higher-resolution image. In this way, the improved higher-resolution image can be used for defect classification. Defect classification using the higher-resolution image provides efficient DOI binning. For example, the higher-resolution image allows for better DOI binning with lower obfuscation points. In this embodiment, if the information determining the sample also includes obfuscation point filtering, then the output of the obfuscation point filtering step can be input into the defect classification. For example, as... Figure 7 As shown, the output 714 of the disturbance point filtering step 710 can be input to the defect classification step 716. In this way, only detection events not filtered out by the disturbance point filtering can be input to the defect classification step.

[0125] Classification of defects based on the higher-resolution images can be performed using any suitable defect classification method or algorithm known in the art, including both non-DL classification methods and DL classification methods. For example, the embodiments described herein can provide the improved defect classification described herein simply by using the currently used defect classifier with a simulated higher-resolution image generated by the trained ML model described herein instead of a lower-resolution test image (or by applying the currently used defect classifier to a simulated higher-resolution image). In this way, the higher-resolution image can be input to the defect classification and used in the same manner as any other input to the defect classification.

[0126] The output from the defect classification step can be input to a post-processing step 718, which may include defect inspection and the generation of final inspection results. Post-processing may include sampling defects not filtered by the scrambling point filter and / or classified by defect classification for defect inspection. Defect inspection can be performed to generate a real, non-simulated, higher-resolution image that can then be used to determine which of the sampled events are actual defects and which are scrambling points, and to determine the defect type or classification of the actual defects. In this way, defect inspection can determine which of the detected events are DOIs and use the non-simulated, higher-resolution image to identify the type or classification of the DOI (when different types of DOIs may exist on the sample).

[0127] In the embodiments described herein, the higher-resolution images generated by the trained ML model can largely eliminate the need for defect inspection, as the defect inspection process required to generate the higher-resolution images typically used for perturbation filtering and defect classification is unnecessary. However, in the embodiments described herein, defect inspection may include actual higher-resolution images of samples generated for events, filtered and unfiltered events, and / or classified or unclassified defects, for several reasons, such as testing the performance of perturbation filtering, verifying defect classification results, and classifying any defects that cannot be classified based on simulated higher-resolution images.

[0128] In some cases, the actual higher-resolution images generated by the defect inspection tool can also be used to verify whether the trained ML model is still being adequately executed and / or whether one or more parameters of the trained ML model should be tuned, which can be executed to illustrate variations or drifts in the process of forming a design on a sample or in the inspection tool.

[0129] The embodiments described herein can significantly reduce the time and cost involved in defect inspection for checking inspection results for scrambling points and / or performing defect classification. The embodiments described herein can even enable inspection-free inspection, where higher-resolution images generated from inspection images by a trained ML model can be used in place of any defect inspection images, thereby reducing or even eliminating the need for defect inspection time on the tool. In other words, the images generated from inspection can be used for defect inspection type functions (such as scrambling point filtering and defect classification) using higher-resolution images generated by a trained ML model, thereby shifting defect inspection to the inspection tool.

[0130] The computer subsystem may be configured to store various information, images, etc., generated by the embodiments described herein. For example, the computer subsystem may be configured to store any or all simulated higher-resolution images of the sample and any or all information determined from the simulated higher-resolution images for the sample. Such images and information may also be stored together with any or all images generated by the testing subsystem for the sample during testing. Any or all of this information may be stored in any suitable manner and in any computer-readable storage medium described herein.

[0131] In another example, the computer subsystem may be configured to store a trained ML model for testing samples or other samples of the same type. The computer subsystem may be configured to store this trained ML model in a scheme, or by generating a scheme for testing that will use the trained ML model. As used herein, the term "scheme" is defined as a set of instructions that can be used by a tool to perform a process on a sample. In this way, generating a scheme may include generating information about how to perform the process, which can then be used to generate instructions for performing the process. The computer subsystem may also store any information (e.g., filename and its storage location) that can be used to identify, access, and / or use the trained ML model. The stored model information may also include the model's procedure code, instructions, algorithms, etc. The model and / or its information may be stored in any suitable manner on any computer-readable storage medium described herein.

[0132] The trained ML model and / or its information may be stored along with any other results described herein and may be stored in any manner known in the art. Storage media may include any storage media described herein or any other suitable storage media known in the art. After storage, the information may be accessed in the storage media and used by any of the methods or system embodiments described herein, formatted for display to a user, used by another software module, method, or system, etc. For example, the embodiments described herein may generate an inspection scheme as described above. The inspection scheme may then be stored and used by a system or method (or another system or method) to inspect samples or other samples, thereby generating information about the samples or other samples (e.g., defect information).

[0133] The results and information generated from testing a sample or other samples of the same type can be used in various ways by the embodiments and / or other systems and methods described herein. Such functionality includes (but is not limited to) modifying processes in a feedback or feedforward manner, such as processes or steps that have been or will be performed on the tested sample or other samples. For example, the computer subsystem described herein can be configured to determine one or more modifications to the process performed on a sample as described herein and / or to the process to be performed on the sample based on detected defects. Modifications to the process may include any suitable changes to one or more parameters of the process. The computer subsystem described herein preferably determines such modifications that can reduce or prevent defects on other samples on which modified processes are performed, correct or eliminate defects on the sample in another process performed on the sample, compensate for defects in another process performed on the sample, etc. The computer subsystem described herein can determine such modifications in any suitable manner known in the art.

[0134] Such changes can then be sent to a semiconductor manufacturing system (not shown) or a computer subsystem and a storage medium (not shown) accessible by the semiconductor manufacturing system. The semiconductor manufacturing system may or may not be part of the system embodiments described herein. For example, the computer subsystem and / or inspection subsystem described herein may be coupled to the semiconductor manufacturing system (e.g., via one or more common elements such as a housing, power supply, sample handling device, or mechanism, etc.). The semiconductor manufacturing system may include any semiconductor manufacturing system known in the art, such as lithography tools, etching tools, chemical mechanical polishing (CMP) tools, deposition tools, and the like.

[0135] Therefore, as described herein, the embodiments can be used to set up new testing procedures or protocols. The embodiments can also be used to modify existing testing procedures or protocols, whether they are for a sample, a testing procedure or protocol established for a sample, or a testing procedure or protocol suitable for another sample.

[0136] The embodiments described herein are not limited to the establishment or modification of inspection schemes or processes. For example, the embodiments described herein can also be used to set up or modify schemes or processes for metrics, defect inspection, etc., in a similar manner. Specifically, the ML model described herein may be pre-trained and retrained depending on a set up or modified process (e.g., generating higher resolution images from images produced during the process). Then, depending on the set up or modified process or scheme, the higher resolution images generated by the trained ML model can be used to perform one or more functions during the process or during post-processing of the output generated by the process. Although many types of quality control processes (such as metrics and defect inspection) are generally configured to produce generally high-resolution images or other outputs of samples, by using the trained ML model described herein to simulate higher resolution images of samples, the sample resolution of the images generated by it during the process can be reduced, which can increase output.

[0137] The various embodiments of the above systems can be combined into a single embodiment.

[0138] Another embodiment relates to a computer-implemented method for determining information about samples. The method includes pre-training, retraining, input, and determination steps as described herein, performed by one or more computer systems. One or more components are performed by one or more computer systems. One or more components include a trained ML model configured to transform images of samples generated by a testing subsystem into higher-resolution images of the samples.

[0139] Each step of the method may be performed as further described herein. The method may also include any other steps that can be performed by the systems, computer systems, components, and / or trained ML models described herein. The computer system may be configured according to any embodiment described herein, such as computer subsystem 102. One or more components and the trained ML model may also be configured according to any embodiment described herein. The method may be performed by any system embodiment described herein.

[0140] Another embodiment relates to a non-transitory computer-readable medium storing program instructions executable on one or more computer systems for performing a computer-implemented method for determining information of a sample. Figure 8 This embodiment is illustrated in the image. Specifically, as shown in the image... Figure 8 As shown herein, the non-transitory computer-readable medium 800 contains program instructions 802 that can be executed on a computer system 804. A computer-implemented method may include any step of any method described herein.

[0141] Program instructions 802 that implement methods such as those described herein may be stored on a computer-readable medium 800. The computer-readable medium may be a storage medium, such as a magnetic disk or optical disk, magnetic tape, or any other suitable non-transitory computer-readable medium known in the art.

[0142] Program instructions can be implemented in any of a variety of ways, including procedural, component-based, and / or object-oriented technologies. For example, program instructions can be implemented using ActiveX controls, C++ objects, JavaBeans, Microsoft Foundation Classes (“MFC”), SSE (Streaming SIMD Extensions), or other technologies or methods, as desired.

[0143] Computer system 804 may be configured according to any of the embodiments described herein.

[0144] Based on this description, those skilled in the art will understand further modifications and alternative embodiments of various aspects of the invention. For example, methods and systems for determining information about samples are provided. Therefore, this description is to be construed as illustrative only and for the purpose of teaching those skilled in the art the general manner of implementing the invention. It should be understood that the forms of the invention shown and described herein are to be considered as presently preferred embodiments. Those skilled in the art, having benefited from the description of the invention, will understand that all elements and materials may replace those described herein, some processes may be interchanged, and certain features of the invention may be utilized independently. Changes may be made to the elements described herein without departing from the spirit and scope of the invention as set forth in the appended claims.

Claims

1. A system configured to determine information about a sample, comprising: The inspection subsystem is configured to generate images of the samples; One or more computer subsystems; and One or more components, executed by the one or more computer subsystems, wherein the one or more components include a trained machine learning model configured to transform the image of the sample generated by the inspection subsystem into a higher resolution image of the sample. The one or more computer subsystems are configured to: An initial machine learning model is pre-trained using a pre-training set, thereby generating a pre-trained machine learning model, wherein the pre-training set includes simulated inspection images of test samples specified only as pre-training inputs and corresponding simulated higher-resolution images of the test samples specified only as pre-training outputs, and wherein the simulated inspection images and the corresponding simulated higher-resolution images include simulated defects. The pre-trained machine learning model is retrained using a training set to generate the trained machine learning model, wherein the training set includes images generated by the verification subsystem for the test samples as training inputs and corresponding higher resolution images of the test samples generated by the high-resolution imaging system as training outputs. The image generated by the inspection subsystem for the sample during inspection is input into the trained machine learning model to generate the higher resolution image of the sample. and Information about the sample is determined from the resulting higher-resolution image.

2. The system according to claim 1, wherein the number of images in the training set is less than the number of simulated test images in the pre-training set, and wherein the number of higher resolution images in the training set is less than the number of simulated higher resolution images in the pre-training set.

3. The system of claim 1, wherein the one or more computer subsystems are further configured to generate the simulated test image and the simulated higher resolution image from the design of the test sample.

4. The system of claim 3, wherein generating the simulated inspection image and the simulated higher resolution image comprises: The process variations of the design formed on the test sample are simulated by applying a perturbation representing the expected process variation to the design to generate a perturbation design.

5. The system of claim 4, wherein a Gaussian process model is used to apply the disturbance to the design.

6. The system of claim 4, wherein generating the simulated inspection image further comprises: The perturbation design is transformed into the image domain.

7. The system of claim 6, wherein a partially coherent model is used to perform the transformation of the perturbation design into the image domain.

8. The system of claim 1, wherein the simulated test image, the simulated higher resolution image, the image generated for the test sample, and the higher resolution image of the test sample comprise only optical images.

9. The system of claim 1, wherein the simulated test image and the image generated for the test sample comprise only optical images, and wherein the simulated higher resolution image and the higher resolution image of the test sample comprise only electron beam images.

10. The system of claim 1, wherein the simulated test image, the simulated higher resolution image, the image generated for the test sample, and the higher resolution image of the test sample comprise only electron beam images.

11. The system of claim 1, wherein the training set further includes the design of the samples.

12. The system of claim 1, wherein the initial machine learning model is a generative network.

13. The system of claim 1, wherein the initial machine learning model is a generative adversarial network.

14. The system of claim 1, wherein the initial machine learning model is an autoencoder.

15. The system of claim 1, wherein the one or more computer subsystems are further configured to input the design of the sample along with the image generated by the testing subsystem for the sample during the testing into the trained machine learning model.

16. The system of claim 1, wherein determining the information of the sample from the generated higher resolution image comprises: Based on the generated higher resolution image, scrambling point filtering of defects detected by the inspection is performed, thereby generating a filtered defect group for the sample.

17. The system of claim 1, wherein determining the information of the sample from the generated higher resolution image comprises: Based on the higher resolution images generated, the defects detected by the test are classified.

18. The system of claim 1, wherein the sample is a wafer.

19. A non-transitory computer-readable medium storing program instructions executable on one or more computer systems to perform a computer-implemented method for determining information of a sample, wherein the computer-implemented method includes: An initial machine learning model is pre-trained using a pre-training set, thereby generating a pre-trained machine learning model, wherein the pre-training set includes simulated inspection images of test samples specified only as pre-training inputs and corresponding simulated higher-resolution images of the test samples specified only as pre-training outputs, and wherein the simulated inspection images and the corresponding simulated higher-resolution images include simulated defects. The pre-trained machine learning model is retrained using a training set to generate a trained machine learning model, wherein the training set includes images generated by the verification subsystem for the test samples as training inputs and corresponding higher resolution images of the test samples generated by the high-resolution imaging system as training outputs. One or more of the components are executed by the one or more computer systems, and the one or more components include the trained machine learning model configured to transform an image of a sample generated by the inspection subsystem into a higher resolution image of the sample. The image generated by the inspection subsystem for the sample during inspection is input into the trained machine learning model to generate the higher resolution image of the sample. and Information about the sample is determined from the resulting higher-resolution image.

20. A computer-implemented method for determining information of a sample, comprising: An initial machine learning model is pre-trained using a pre-training set, thereby generating a pre-trained machine learning model, wherein the pre-training set includes simulated inspection images of test samples specified only as pre-training inputs and corresponding simulated higher-resolution images of the test samples specified only as pre-training outputs, and wherein the simulated inspection images and the corresponding simulated higher-resolution images include simulated defects. The pre-trained machine learning model is retrained using a training set to generate a trained machine learning model, wherein the training set includes images generated by the verification subsystem for the test samples as training inputs and corresponding higher resolution images of the test samples generated by the high-resolution imaging system as training outputs. One or more of the components are executed by one or more computer systems, and the one or more components include the trained machine learning model configured to transform images of samples generated by the inspection subsystem into higher resolution images of the samples. The image generated by the inspection subsystem for the sample during inspection is input into the trained machine learning model to generate the higher resolution image of the sample. and Information about the sample is determined from the resulting higher resolution image, wherein the pre-training, retraining, input, and determination are performed by the one or more computer systems.

Images