Defect synthesis and detection via defect generation pre-trained transformers for semiconductor applications

Abstract

Methods and systems for determining information of a sample are provided. A system includes a computer subsystem and one or more components executed by the computer subsystem. The one or more components include a pre-trained defect generation pre-trained transformer (DefectGPT) encoder configured for determining information of a sample based on one or more inputs specific to the sample. The computer subsystem is configured for inputting the one or more inputs into the pre-trained DefectGPT encoder.

Description

Technical Field

[0001] This invention generally relates to methods and systems for determining information about a sample. Some embodiments relate to methods and systems for defect synthesis and detection via a defect generation pre-trained converter for semiconductor applications. Background Technology

[0002] The following descriptions and examples are not considered prior art simply 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 manufacturing processes to form various features and multiple layers of the semiconductor device. For example, photolithography is a semiconductor manufacturing process involving transferring a pattern from a photomask to a photoresist disposed on a semiconductor wafer. Additional examples of semiconductor manufacturing 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 in some arrangement and then separated into individual semiconductor devices.

[0004] Inspection processes are used at every step of the semiconductor manufacturing process to detect defects in samples, driving higher yields and therefore 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, because even small defects can lead to device failure.

[0005] Defect re-inspection typically involves re-inspecting defects detected as defects by the inspection process and using a high-magnification optical system or scanning electron microscope (SEM) to generate additional information about the defects at higher resolution. Therefore, defect re-inspection is performed at discrete locations on the sample where defects were previously detected by inspection. The higher-resolution data of defects generated through defect re-inspection is better suited for determining defect properties such as profile, roughness, and more accurate size information. Compared to inspection, defects can often be more accurately classified into defect types based on the information determined through defect re-inspection.

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

[0007] The metrological process differs from the defect re-inspection process in that, unlike the defect re-inspection process which revisits defects detected by inspection, the metrological process can be performed at locations where defects were not detected. In other words, unlike defect re-inspection, the location for performing the metrological process on the sample can be independent of the results of the inspection process performed on the sample. Specifically, the location for performing the metrological process can be selected independently of the inspection results. Furthermore, because the location for performing the metrological process on the sample can be selected independently of the inspection results, unlike defect re-inspection where the location for defect re-inspection on the sample cannot be determined until the sample's inspection results are generated and available, the location for performing the metrological process can be determined before the inspection process has been performed on the sample.

[0008] Even after establishing suitable hardware parameters for the yield-related processes described above, the methods, systems, and algorithms used to determine sample information from tool outputs can still be challenging. For example, even the best image of a sample generated by the best available tool may not be good enough to determine sample information, and some image processing may be required. At other times, it may be difficult to determine a suitable method for determining sample information from sample images or other tool outputs.

[0009] Therefore, numerous methods, algorithms, systems, etc., have been created for determining sample information in the process described above. In addition to many other currently used image processing techniques, some examples of currently used defect detection methods include computational reference type methods and algorithms available on some currently available tools (e.g., the Multi-Die Automatic Thresholding (MDAT) algorithm). Further examples of currently used defect detection methods include active learning for defect classifier training, learnable defect detection for semiconductor applications, and unsupervised or self-supervised deep learning (DL) for semiconductor-based applications.

[0010] While many of the defect detection methods described above have achieved at least some success in their implementations, many have at least some minor drawbacks or areas for improvement. For example, computational reference methods can be based on linear combinations and cannot approximate higher-order (imaging / signal) information, which is crucial for distinguishing weak defect signals from moderate / weak noise. In another instance, active learning is advantageous for iteratively training DL defect classifiers or detectors. However, for these methods to work, sufficient training samples are needed between each training iteration; and the model must be trained per semiconductor layer or per process step. Furthermore, DL-based or convolutional neural network (CNN)-based detectors typically must be trained per semiconductor layer or per process step. Similarly, unsupervised or self-supervised DL methods must be trained per semiconductor layer or per process step.

[0011] Even if a suitable method for processing tool output is found, the challenge of generating suitable training data may still exist. In one such example, to generate a suitable defect detection method, it is typically necessary to select a relatively large number of defect instances and then train the defect detection method. These defect instances may be difficult to obtain or generate for several reasons, such as the relatively limited number of defects on any given sample, the difficulty and time involved in identifying such defects, and even the time and cost of generating tool output for such defects.

[0012] Therefore, methods and systems have been developed to synthesize defects that can be used in building processes (such as those described above). Some of these methods use machine learning / deep learning (ML / DL) generation models to generate defect-free optical images from design information for a single optical pattern. Other ML / DL generation models have been created to generate defect-free optical images from design information for multiple optical patterns. ML / DL generation models have also been created to generate defect-free optical images from optical images (collected under a given pattern) for multiple optical patterns. In some cases, computer vision algorithms (e.g., painting, blending, etc.) can produce relatively low-quality defect images. Additionally, optical simulation can be used to model defect images.

[0013] However, currently available defect synthesis methods have several significant drawbacks. For example, an ML / DL generation model created to produce a defect-free optical image from design information for a single optical mode can only produce a background integrated circuit (IC) pattern for that single optical mode, and not a defect. In another instance, an ML / DL generation model that can produce a defect-free optical image from design information for multiple optical modes can only produce a background IC pattern for multiple optical modes, not a defect within the context of multiple optical modes. In yet another instance, an ML / DL generation model created to produce a defect-free optical image from an optical image (collected in a given mode) for multiple optical modes can only transfer the background IC pattern from one optical mode to other optical modes, not a defect within the context of other optical modes. Computer vision algorithms can be disadvantageous because the generated defects have relatively low quality, inconsistent with the properties of real defects, including shape, intensity, diffusion, etc. Furthermore, it is difficult to incorporate process variations, conditions affecting pattern variations, and system and random noise into the optical simulation of the defect image. Additionally, these methods are computationally intensive and therefore impractical for synthesizing a relatively large number of defects.

[0014] Therefore, it would be advantageous to develop systems and methods for determining information about a sample that do not have one or more of the disadvantages described above. Summary of the Invention

[0015] The following description of various embodiments should not be construed in any way as limiting the subject matter of the appended claims.

[0016] One embodiment relates to a system configured to determine information about a sample. The system includes a computer subsystem and one or more components executed by the computer subsystem. The one or more components include a pre-trained DefectGPT encoder, the pre-trained DefectGPT encoder being configured to determine information about the sample based on one or more sample-specific inputs. The computer subsystem is configured to input the one or more inputs into the pre-trained DefectGPT encoder. The system may be further configured as described herein.

[0017] Another embodiment relates to a computer-implemented method for determining information about a sample. The method includes inputting one or more sample-specific inputs into a pre-trained DefectGPT encoder, the pre-trained DefectGPT encoder being configured to determine information about the sample based on the one or more inputs. 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.

[0018] Another embodiment relates to a non-transitory computer-readable medium storing program instructions executable on a computer system to perform a computer-implemented method for determining information about 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 for which the program instructions are executable may include any other steps of any other method described herein. Attached Figure Description

[0019] Further advantages of the invention will become apparent to those skilled in the art upon the following detailed description of preferred embodiments and upon reference to the accompanying drawings, wherein:

[0020] Figure 1 and 2 This is a schematic diagram of a side view illustrating an embodiment of a system configured as described herein;

[0021] Figure 3 This is a flowchart illustrating an embodiment of a pre-training defect generating a pre-trained transformer (DefectGPT) model;

[0022] Figure 4 and 5 This is a flowchart illustrating an embodiment of defect detection performed via defect hints using a pre-trained DefectGPT encoder;

[0023] Figure 6 This is a block diagram illustrating an embodiment of a network A configured to generate defect signals;

[0024] Figure 7 This is a block diagram illustrating an embodiment of a network B configured to generate a flawless analog image;

[0025] Figure 8 and 9 These are block diagrams illustrating embodiments of networks C1 and C2 configured to generate simulated images with defects;

[0026] Figure 10 This is a block diagram illustrating an embodiment of a network D configured to generate a simulated image with defects using a diffusion model; and

[0027] Figure 11 This is a block diagram illustrating one embodiment of a non-transitory computer-readable medium storing program instructions for causing a computer system to perform the computer-implemented methods described herein.

[0028] While the invention is susceptible to various modifications and alternatives, specific embodiments thereof are shown by way of example in the drawings and will be described in detail herein. The drawings may be drawn not 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

[0029] Turning now to the figures, it should be noted that the figures are not drawn to scale. In particular, the scale of some elements in the figures is greatly exaggerated to emphasize the characteristics of the elements. It should also be noted that the figures are not drawn to the same scale. The same element symbols have been used to indicate elements that can be similarly configured and shown in more than one figure. Unless otherwise mentioned herein, any element described and shown may include any suitable commercially available element.

[0030] Generally, the embodiments described herein are configured to determine sample information via defect synthesis and / or defect detection. Defect synthesis is typically defined herein as the process of simulating defects of interest (DOI) in semiconductor applications to understand process conditions, process windows, and achieve yield improvements in semiconductor manufacturing. The embodiments described herein introduce a technique called Defect Generation Pre-trained Transformer (DefectGPT) for AI-enabled defect synthesis processes, which enables synthesized defect images for optical inspectors to be well aligned with the process conditions, process variations, and defect distribution of a given process wafer. The synthesized defect images can be applied to defect detection in inspection and metrology applications. Furthermore, the defect detection and defect synthesis methods and systems described herein are all based on DefectGPT models constructed using different training and inference methods.

[0031] 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 samples that can be used with them. For example, the embodiments described herein can be used with, for example, photomasks, flat panels, personal computer (PC) boards, and other semiconductor samples.

[0032] Figure 1 An embodiment of a system configured to determine information about a sample is shown. In some embodiments, system 10 includes an inspection subsystem, such as inspection subsystem 100. The inspection subsystem includes and / or is coupled to a computer subsystem, such as computer subsystem 36 and / or one or more computer systems 102.

[0033] Generally, the testing subsystem described herein includes at least an energy source and a detector. The energy source is configured to generate energy directed to the sample. The detector is configured to detect the energy from the sample and generate an output in response to the detected energy.

[0034] In a light-based inspection subsystem, the energy guided to the sample contains light, and the energy detected from the sample also contains light. For example, as... Figure 1 As shown, the inspection subsystem includes an illumination subsystem configured to direct light to sample 14. The illumination subsystem includes at least one light source, such as light source 16. The illumination subsystem is configured to direct 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 at an angle of incidence through optical element 18 and then lens 20 to sample 14. 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 to be detected on the sample.

[0035] The illumination subsystem can be configured to direct light to the sample at different incident angles. For example, the testing subsystem can be configured to change one or more parameters 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 one example, the inspection subsystem can be configured to move the light source 16, optical element 18, and lens 20 such that light is directed to the sample at different angles of incidence or normal (or near-normal) incident angles. The illumination subsystem can have any other suitable configuration known in the art for sequentially or simultaneously directing light to the sample at one or more incident angles.

[0036] The illumination subsystem can also be configured to direct light with different characteristics to the sample. For example, 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 replacing one spectral filter with another) so that light of different wavelengths can be directed to the sample at different times.

[0037] Light source 16 may comprise a broadband plasma (BBP) source. In this way, 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 any suitable laser known in the art configured to generate light of any suitable wavelength. Additionally, the laser may be configured to generate monochromatic or near-monochromatic light. In this way, the laser may be a narrowband laser. The source may also comprise a multicolor source that generates light of multiple discrete wavelengths or bands.

[0038] Light from the optical element 18 can be focused onto the sample 14 through lens 20. Although lens 20... Figure 1 While shown as a single-refractive optical element, lens 20 may actually comprise several refractive and / or reflective optical elements that combine and focus light from the optical element onto the sample. Figure 1 The 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, polarizing components, spectral filters, spatial filters, reflective optics, apodizers, beam splitters, apertures, and the like, which may include 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 to be tested.

[0039] The inspection subsystem may also include a scanning subsystem configured to change the position of light on the sample to which it is directed and from which light is detected, and may result in 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. Alternatively or additionally, the inspection subsystem may be configured such that one or more optical elements of the inspection subsystem perform a certain scan of the sample, such that light can be directed to different positions on the sample and detected from different positions. The light may scan across the sample in any suitable manner (e.g., in a serpentine or helical path).

[0040] The inspection subsystem further includes one or more detection channels. At least one of the detection channels includes a detector configured to detect light originating from the sample due to illumination of the sample by the system, and to generate an output in response to the detected light. Figure 1 The inspection subsystem shown includes two detection channels: one formed by a light collector 24, element 26, and detector 28, and the other formed by a light collector 30, element 32, and detector 34. The 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 channels are 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).

[0041] exist Figure 1In this embodiment, two detection channels are shown positioned in the plane of the paper, and the illumination subsystem is also shown positioned in the plane of the paper. Therefore, in this embodiment, the two detection channels are positioned in the plane of incidence (e.g., centered in the plane of incidence). However, one or more of the detection channels may be positioned outside the plane of incidence. For example, the detection channel formed by the light collector 30, element 32, and detector 34 may be configured to collect and detect light scattered from the plane of incidence. Therefore, this detection channel may be collectively referred to as a "side" channel, and this side channel may be centered in a plane substantially perpendicular to the plane of incidence.

[0042] 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., only one detection channel or two or more detection channels). The detection channel formed by the light 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 and positioned on the opposite side of the incident plane. Therefore, the inspection subsystem may include a detection channel comprising the light collector 24, element 26, and detector 28, centered in the incident plane, and configured to collect and detect light at a scattering angle normal to or near normal to the sample surface. Therefore, this detection channel may be collectively referred to as the “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., one top channel and two side channels), and each of the at least three channels is configured to collect light at a scattering angle different from that of each of the other light collectors.

[0043] As further described above, one or more of the detection channels can be configured to detect scattered light. Therefore, Figure 1 The inspection subsystem shown herein can be configured for dark field (DF) inspection. However, the inspection subsystem may also, or alternatively, include a detection channel configured for bright field (BF) inspection. Therefore, the inspection subsystem described herein can be configured for DF-only inspection, BF-only inspection, or both DF and BF inspection. Although each of the light collectors... Figure 1 It is shown as a single refractive optical element, but each of the light collectors may contain refractive optical elements and / or reflective optical elements.

[0044] 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 (e.g., intensity) of the scattered light, but may not be configured to detect such characteristics that vary depending on position within the imaging plane. Therefore, the output generated by each of the detectors in each of the detection channels may be a signal or data, rather than an image signal or image data. In such examples, the computer 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 configured to generate an imaging signal or image data. Therefore, the inspection subsystem may be configured to generate images in several ways.

[0045] Computer subsystem 36 may be coupled to the detector of the verification subsystem in any suitable manner (e.g., via one or more transmission media, which 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’s output, as further described herein. Computer subsystem 36 may be further configured as described herein.

[0046] Computer subsystem 36 (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.

[0047] If the system contains more than one computer subsystem, then the different computer subsystems can be coupled to each other, enabling the transmission of images, data, information, instructions, etc., between the computer subsystems. For example, computer subsystem 36 can be coupled to computer system 102 via any suitable transmission medium, such as by... Figure 1 The dashed lines in the diagram indicate that the transmission medium may include any suitable wired and / or wireless transmission medium known in the art. Two or more of these computer subsystems may also be effectively coupled by a shared computer-readable storage medium (not shown).

[0048] In the electron beam inspection subsystem, the energy guided to the sample contains electrons, and the energy detected from the sample also contains electrons. Figure 2 In one 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 can be configured as described above. Alternatively, this inspection subsystem can be identical to that described above. Figure 1 The method shown in the diagram is coupled to one or more computer subsystems.

[0049] Also Figure 2 As shown, the electron column includes an electron beam source 126 configured to generate electrons focused onto the 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 current selection aperture, an objective lens, and a scanning subsystem, all of which may include any such suitable elements known in the art.

[0050] 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, which may be the same scanning subsystem included in element 130.

[0051] The electron column may include any other suitable element known in the art. Furthermore, the electron column may be further configured as described in the following patents: U.S. Patent No. 8,664,594, issued April 4, 2014, to Jiang et al.; U.S. Patent No. 8,692,204, issued April 8, 2014, to Kojima et al.; U.S. Patent No. 8,698,093, issued April 15, 2014, to Gubbens et al.; and U.S. Patent No. 8,716,662, issued May 6, 2014, to MacDonald et al., all of which are incorporated herein by reference in their entirety.

[0052] Despite the electron column Figure 2 The electron beam is shown as being configured such that electrons are guided to and scattered from the sample at an angle of incidence at a tilted 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 produce sample output using multiple modes (e.g., using different illumination angles, collection angles, etc.), as further described herein. The multiple modes of the electron beam inspection subsystem can differ in terms of the parameters produced by any output of the inspection subsystem.

[0053] Computer subsystem 124 may be coupled to detector 134, as described above. The detector detects electrons returning from the surface of the sample, 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 perform any of the steps described herein. Figure 2 The inspection subsystem shown in the document can be further configured as described herein.

[0054] This article provides Figure 1 and 2 The configuration of the inspection subsystems that may be included in the system embodiments described herein is explained in general terms. Clearly, the configuration of the inspection subsystems described herein can be modified to optimize the performance of the inspection subsystems, as is typically done when designing commercial inspection systems. Additionally, the systems described herein can be implemented using existing inspection subsystems (e.g., by adding the functionality described herein to an existing inspection system), for example, tools commercially available from KLA Corp., Milpitas, California. For some such systems, the methods described herein may be provided as optional functionality of the inspection system (e.g., in addition to other functionality of the inspection system). Alternatively, the inspection systems described herein can be designed “from scratch” to provide entirely new inspection systems.

[0055] Although the inspection subsystem is described above as a light or electron beam inspection subsystem, it can also be an ion beam inspection subsystem. This inspection subsystem can be as follows: Figure 2 The configuration shown is different only in that the electron beam source can be replaced with any suitable ion beam source known in the field. Additionally, the inspection subsystem can include any other suitable ion beam subsystem, such as those found in commercially available focused ion beam (FIB) systems, helium ion microscopy (HIM) systems, and secondary ion mass spectrometry (SIMS) systems.

[0056] The testing subsystem can be configured to generate sample outputs, such as images, using multiple modes. Generally, a "mode" is defined by the parameter values ​​(or the outputs used to generate images of the sample) of the testing subsystem. Therefore, modes can differ in the value of at least one parameter of the testing subsystem (except for the location where the output is generated on the sample). For example, modes can differ in any one or more modifiable parameters of the testing subsystem (e.g., illumination polarization, angle, wavelength, etc.; detection polarization, angle, wavelength, etc.). The testing subsystem can be configured to scan samples using different modes in the same or different scans, depending on the ability to scan samples simultaneously using multiple modes.

[0057] Similarly, an electron beam subsystem can be configured to generate images using two or more modes, which can be defined by parameter values ​​of the electron beam subsystem used to generate images of the sample. Therefore, the modes can differ in the value of at least one electron beam parameter of the electron beam subsystem. For example, different modes can use different incident angles for illumination.

[0058] In another embodiment, the system includes a metering subsystem. In yet another embodiment, the system includes a defect re-detection subsystem. For example, Figure 1 and 2 The embodiment of the inspection subsystem shown can be modified on one or more parameters to provide different imaging capabilities depending on the application it will be used for. In one such example, if the inspection subsystem is intended for metrology rather than inspection, it can be configured to have a higher resolution. In other words, Figure 1 and 2 The embodiments of the inspection subsystem shown in the description are general and various configurations of an imaging subsystem that can be understood by those skilled in the art to produce systems with different imaging capabilities that are more or less suitable for different applications.

[0059] In this way, the imaging subsystem can be configured to generate outputs suitable for re-inspecting defects on a sample in the case of a defect re-inspection system and suitable for measuring one or more characteristics of the sample in the case of a metrology system. In an embodiment of a defect re-inspection system, Figure 2 The computer subsystem 124 shown can be configured to re-detect defects on sample 128 by applying a defect re-detection method to the output generated by detector 134, and may use the output generated by the detector to determine additional information about the re-detected defects. In a metrology system embodiment, Figure 1 The computer subsystem 36 shown can be configured to determine one or more characteristics of sample 14 using the output generated by detector 28 and / or 34.

[0060] As mentioned above, the testing subsystem is configured to scan a physical version of the sample with energy (e.g., light, electrons, etc.), thereby producing an output of the physical version of the sample. In this way, the testing subsystem can be configured as a "real" subsystem, rather than a "virtual" subsystem. However, the storage media (not shown) and Figure 1The computer subsystem 102 shown herein can be configured as a “virtual” system. Specifically, the storage media and computer subsystem can be configured as a “virtual” verification system, as described in the following commonly assigned patents: U.S. Patent No. 8,126,255 to Bhaskar et al., issued February 28, 2012, and U.S. Patent No. 9,222,895 to Duffy et al., issued December 29, 2015, which are incorporated herein by reference as if described in their entirety. The embodiments described herein can be further configured as described in those patents.

[0061] The system comprises a computer subsystem (which may include the computer subsystems described above or any configuration of any of them) and one or more components executed by the computer subsystems. For example, such as Figure 1 As shown, the system may include a computer subsystem 36 and one or more components 104 executed by the computer subsystem. The one or more components include a pre-trained DefectGPT encoder configured to determine information about a sample based on one or more sample-specific inputs. The computer subsystem is configured to input one or more inputs into the pre-trained DefectGPT encoder, which can be performed in any suitable manner known in the art.

[0062] The inputs can include a variety of different inputs, as further described herein, which is an advantage of the embodiments described herein. In some embodiments, the computer subsystem may generate one or more of the inputs, but more generally, the computer subsystem may obtain one or more of the inputs, for example, from test recipes, from storage media, from one or more other systems and / or methods, from users, etc. For example, inputs such as DOI location prompts and DOI descriptors may be obtained from user input or from test recipes and the like that previously produced test results. Inputs such as optical patterns and process parameters may be obtained from test recipes and process recipes stored in a storage medium, respectively. Other inputs, such as design images, may be obtained from storage media that store designs and / or design images generated from design data. Additional inputs described herein may be obtained in a similar manner.

[0063] In one embodiment, the computer subsystem is configured to detect defects on a sample based on information determined for the sample by a pre-trained DefectGPT encoder. For example, in some embodiments described herein, the pre-trained DefectGPT encoder can be used to detect defects on a sample. In other embodiments, the pre-trained DefectGPT encoder may be configured to generate synthetic information about the sample or defects that may be present on the sample.

[0064] In another embodiment, the computer subsystem is configured to pre-train an initial DefectGPT encoder, thereby generating a pre-trained DefectGPT encoder, and performs pre-training using a training dataset containing sample-specific images and images unrelated to information determining the sample based on one or more inputs. The unrelated images are optional. For example, in another embodiment, the computer subsystem is configured to pre-train an initial DefectGPT encoder, thereby generating a pre-trained DefectGPT encoder, and performs pre-training using a training dataset that includes only sample-specific images. In defect detection constructions, for example, as... Figure 3 As shown, the pre-trained DefectGPT model may involve inputting a training dataset 300 into the DefectGPT encoder 302. The training dataset may contain only the target semiconductor image (i.e., a sample-specific image) or the target semiconductor image and optional out-of-domain images (i.e., images unrelated to determining the sample information). In one such example, when the application is scanning electron microscopy (SEM) defect detection, the target semiconductor image is an SEM image, and the optional out-of-domain images may include other semiconductor images, natural images, or images from other application domains.

[0065] In an additional embodiment, the computer subsystem is configured to pre-train an initial DefectGPT encoder, thereby generating a pre-trained DefectGPT encoder, and the initial DefectGPT encoder is configured to encode the input image into a visual lexical embedding 304. For example, the DefectGPT encoder receives an input image having dimensions (height, width, channels) and encodes it into a visual lexical embedding having a size (H / s, W / s, embedding size), where s is a stride in the range [1, min(H,W)].

[0066] In another embodiment, the computer subsystem is configured to pre-train an initial DefectGPT encoder in a self-supervised manner, thereby generating a pre-trained DefectGPT encoder. For example, a contrastive loss 306 or its equivalent can be used to pre-train the model in a self-supervised manner.

[0067] Next, defect detection can be performed using defect prompts. In some embodiments, a pre-trained DefectGPT encoder is configured to determine the DOI query embedding based on one or more input DOI prompt images. Figure 4 As shown in the example, a pre-trained DefectGPT encoder 402 is configured to determine a DOI query embedding 404 based on a DOI prompt image 400. The DOI prompt image may be a DOI containing defects or partial defects and having a size (N). s, N The DOI hint images are plotted images of type N, where N is any integer (>=1), and s is the stride used in pre-training. DOI hint images can be obtained using any suitable method known in the domain, such as using real-world methods or known good defect detection methods. When more than one input DOI hint image exists for generating query embeddings, a set of query embeddings can be generated using each of the input DOI hint images, and then a combined query embedding can be computed by averaging or weighted averaging them. In this way, combined query images can be generated using different input DOI hint images for different examples of the same type of DOI or for different examples of the same DOI. Generating combined query embeddings in this way increases the robustness of the pre-trained DefectGPT encoder.

[0068] In one embodiment, one or more inputs include a target image for defect detection, determination information includes determining the visual lexical embedding of the target image, and a computer subsystem is configured to determine a similarity metric between the DOI query embedding and the visual lexical embedding, and to determine whether a defect exists in the target image based on the determined similarity metric. For example, such as Figure 5 As shown, any image 500 (which can be any target image of any size for defect detection) can be input into a pre-trained DefectGPT encoder 502, which generates a visual word embedding 504. The computer subsystem then determines a similarity measure between the target visual word embedding 504 and the cue DOI query embedding 404, and generates a detection map 506 via similarity, which may contain information about any location on the sample or in the image where the similarity measure between the visual word embedding and the DOI query embedding is above a predetermined threshold.

[0069] Similarity measures can include any qualitative or quantitative measure that can be used to express how similar different embeddings are, and can be determined by a computer subsystem in any suitable manner known in the relevant domain. A predetermined threshold for separating sufficiently similar embeddings from insufficiently similar embeddings can be determined in any suitable manner (e.g., based on user input).

[0070] The computer subsystem can also generate other inspection results from the inspection map (e.g., defect list 508), which may contain any information on the sample or in the image that has been identified as the location of the defect by a similarity metric.

[0071] Therefore, the embodiments described herein can use visual lexical embeddings and cue DOI query embeddings for defect detection. In this way, the defect detection described herein is performed in a latent space rather than an image space. In other words, unlike most other defect detection methods that involve comparing the actual images themselves (e.g., to generate differential images for subsequent defect testing), the embodiments described herein determine embeddings from different images and then compare the embeddings, rather than the images. Performing defect detection in this manner will likely make the process simpler and more accurate because the embeddings can be determined substantially accurately through the embodiments described herein, and comparing embeddings rather than images is simpler and should produce more accurate results.

[0072] The embodiments described herein are also advantageous because, by determining the cue DOI query embedding, as described herein, defect detection can be used only to find similar DOIs on the test sample. In other words, this defect detection can avoid looking for scrambled points and other non-interesting defects, thereby increasing the accuracy of the inspection and reducing (or even completely eliminating) the scrambled point filtering that can still be performed on these results compared to the normally performed scrambled point filtering.

[0073] Although an embodiment of a cue DOI query embedding is described herein, embodiments may generate multiple cue DOI query embeddings (each of multiple different DOIs having one cue DOI query embedding), and any visual lemma embedding may be compared with the multiple cue DOI query embeddings to see if any of the DOIs is present in the image. Furthermore, although an embodiment of a positive cue query embedding is described herein, cue query embeddings may also be generated and used to identify non-interesting defects and / or obstructions or noise on a sample. In other words, cue query embeddings may be generated for either or both positive and negative instances of an event that may be present on a sample, even though, generally, the user will only be interested in positive instances of the event (e.g., DOI).

[0074] Generally, the defect synthesis embodiments presented herein can be described herein and in Figures 6 to 10 The network is constructed using one or more main steps, illustrated and named networks A, B, C1, C2, and D respectively. All encoders in networks A, B, C1, C2, and D use a pre-trained DefectGPT encoder. Additionally, in these embodiments, one or more components include a decoder configured to generate synthetic defect information from information determined by the pre-trained DefectGPT encoder for the sample. The decoder can be configured as further described herein.

[0075] Figure 6 The network A shown is configured to generate defect signal images for networks C1 and / or C2. For example... Figure 6As shown, in some embodiments, a computer subsystem is configured to generate fused DOI features 604 from DOI location cues 600 and DOI descriptors 602. One or more inputs to a pre-trained DefectGPT encoder 606 contain the fused DOI features, the determined information includes DOI embeddings 608, and one or more components include a decoder 610 configured to generate a defect signal image 612 from the DOI embeddings. In this manner, inputs containing DOI location cues and DOI descriptors (inspection optical parameters, process-related parameters, metrological parameters, etc.) are fused into DOI features. The fused DOI features are then converted into latent DOI features after processing by the pre-trained DefectGPT encoder 606. The defect signal is then generated by the decoder.

[0076] Inspection optical parameters may include the optical conditions of the inspection tool. For example, on some inspection tools, these parameters may include optical wavelength bands, focusing conditions, aperture, etc. Each of these can be a single input parameter, such as wavelength = 200 nm, focus = 0 nm, aperture = BF, etc. Process-related parameters are the process parameters of the process tools used to manufacture the wafer, such as lithography focusing and exposure, etching tool conditions, etc. Metrological parameters may include parameters measured by metrology tools in the current or previous process steps of critical wafer metrology measurements, such as film thickness.

[0077] The fusion performed in step 604 can be performed in a similar manner to other networks described herein (although with different inputs). The DOI location hint 600 can be a spatial matrix of size (height, width, 1), where each element has a value of 0 or 1, where 0 represents a non-DOI location and 1 represents a DOI location. The optical parameters in DOI descriptor 602 are spatial matrices with the same H and W but potentially different numbers of channels (i.e., size (H, W, C)). The C value represents the number of optical parameters. For example, if we consider three optical parameters, wavelength, aperture, and focus, then the matrix would be (H, W, 3). All elements in a parameter channel will have the same encoded value to indicate the parameter selection. A similar construction can be used for process-related parameters in DOI descriptor 602 with a matrix size of (H, W, C'), where C' is the number of process parameters. Metrological parameters are spatial matrices of size (H, W, C''). A spatial matrix (number of layers). Each element in the matrix represents the measurement of the C''-th parameter at spatial locations H and W at the selected layer. The fusion operation can potentially combine four matrices into a single matrix through concatenation:

[0078] concat: (H, W, 1) + (H, W, C) + (H, W, C') + (H, W, C'' n) =>(H, W, 1+C+C'+C'' n)

[0079] Network B is configured to generate defect-free simulated images for both networks C1 and C2. For example... Figure 7 As shown in another embodiment, a computer subsystem is configured to generate fused patterned features 704 from a design image 700 of the sample, optical pattern information, and process parameter information (optical pattern, process parameters 702), to one or more inputs of a pre-trained DefectGPT encoder 706 containing fused patterned features. The determined information includes pattern embedding 708, and one or more components include a decoder 710 configured to generate a simulated image 712 of the sample without defects from the pattern embedding. In this way, the input containing the design image, optical pattern, and process parameters is fused into the patterned features, which can be performed as further described herein. The patterned features are converted into pattern latent features after being processed by the pre-trained DefectGPT encoder 706. The decoder generates a simulated image without defects.

[0080] Network C1 is configured to generate simulated images with defects. For example... Figure 8 As shown, in another embodiment, a computer subsystem is configured to generate a fused input (via a fusion operation 804) from a defect signal image 800 of the sample and a simulated image 802 of the sample without defects. One or more inputs to a pre-trained DefectGPT encoder 806 contain the fused input (generated by the fusion operation 804), the determined information includes a simulated embedding 808, and one or more components include a decoder 810 configured to generate a simulated image 812 of the sample with defects from the simulated embedding. In this manner, the input containing the defect signal (generated by network A) and the simulated image without defects (from network B) is fused. The fused features are converted into simulated latent features after processing by the pre-trained DefectGPT encoder 806. The simulated image with defects is generated by the decoder.

[0081] Network C2 is also configured to generate flawed simulated images. For example... Figure 9As shown, in one embodiment, the determined information includes a DOI embedding 900 (from network A), a pattern embedding 902 (from network B), and a combined embedding 904 generated from the DOI embedding and the pattern embedding. One or more components include a decoder 906 configured to generate a simulated image 908 of defects on the sample from the combined embedding. In this embodiment, the input, including latent features of the DOI (e.g., generated by a pre-trained DefectGPT encoder 606 in network A) and latent features of the pattern (e.g., generated by a pre-trained DefectGPT encoder 706 in network B), is then combined. The decoder then generates the simulated image of defects.

[0082] Optimization of the diffusion process / model is constructed in Figure 10 The network shown is named Network D. Network D is configured to generate a simulated image with defects. In this embodiment, one or more inputs include a design image 1010 of the sample, and one or more components include a first diffusion process model 1016, which is configured to generate a design embedding 1020 from information determined for the sample by a pre-trained DefectGPT encoder 1012. In this way, the design image is processed by the pre-trained DefectGPT encoder 1012.

[0083] In one embodiment, one or more components include a conditional encoder 1008 configured to determine additional information about the sample from input 1000, which includes DOI location hints 1002, DOI descriptors 1004, optical mode information, and process parameter information (optical mode / process parameters 1006). One or more components also include a second diffusion process model 1014 configured to generate a conditional embedding 1018 from the additional information. In this manner, the conditional encoder 1008 processes the input 1000, which includes DOI location hints and DOI descriptors (verification optical parameters, process-related parameters, metrological parameters, etc.), as well as the optical mode and process parameters. Furthermore, the diffusion process / model follows both the conditional encoder and the pre-trained DefectGPT encoder. The diffusion process / model can have any suitable configuration known in the art.

[0084] In another embodiment, one or more components include a decoder 1022 configured to generate a flawed simulated image 1024 of the sample based on design embedding 1020 and conditional embedding 1018. In this manner, the decoder generates the flawed simulated image.

[0085] In one embodiment, the pre-trained DefectGPT encoder is configured as a visual transformer (ViT) or a Swin transformer. Alternatively, the pre-trained DefectGPT encoder can be any transformer-based model, including but not limited to transformers, ViTs, hybrid convolutional neural networks (CNNs) with transformers / ViTs, and Swin transformers. ViTs can be configured as described by Dosovitskiy et al. in “An Image isWorth 16x16 Words: Transformer for Image Recognition at Scale,” a conference paper published at ICLR 2021, arXiv:2010.11929v2, June 3, 2021, page 22, which is incorporated herein by reference as if fully described. The Swin transform can be configured as described by Liu et al. in “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”, arXiv:2103.14030v2, August 17, 2021, page 14, which is incorporated herein by reference as if described in its entirety. The pre-trained DefectGPT encoder described herein can be further configured as described in those references.

[0086] In another embodiment, one or more components include a decoder configured to determine additional information about a sample from information determined for that sample, and the decoder is configured as a ViT or Swin transform. Additionally, the DefectGPT decoder model can be any transform-based model, including, but not limited to, transforms, ViT, hybrid CNNs with transforms / ViT, CNNs, and Swin transforms. The decoder described herein can also be further configured as described in the references incorporated above.

[0087] The process of establishing the embodiments described herein can be performed in various ways as described herein. The training data input for network A may include DOI location hints and DOI descriptors. DOI location hints may be binary graphs representing defect locations. DOI descriptors contain characteristic attributes of the DOI (optical, manufacturing, tooling conditions, etc.). The live data used for network A may be defects generated by a physical model or actual defects.

[0088] The training data input for Network B includes a design image, optical patterns, and process parameters. The design image represents the background pattern of the optical image. The optical patterns and process parameters are used to represent different optical patterns and process parameters. The live data used for Network B is a real optical image.

[0089] The training data input for network C1 includes defect signals generated by network A and simulated images without defects generated by network B. The real-world data used for network C1 can be defective images generated by a physical model or real optical images with defects.

[0090] The training data input for network C2 includes latent DOI features generated by network A and pattern latent features generated by network B. The live data used for network C2 can be a defective image generated by a physical model or a defective real optical image.

[0091] The training steps may include initializing the encoder backbone from scratch or the parameters of a pre-trained ImageNet DefectGPT encoder. (“Backbone” refers to the network architecture in the DefectGPT encoder.) The best way to do this is through self-supervised learning, where the encoder backbone is pre-trained on a relatively large cluster of wafer datasets. The encoder backbone may also be jointly learned with a defect detection model. The parameters of the decoder backbone may be initialized from scratch. In a general construct, the model described herein may be trained as follows: (1) Network A; (2) Network B; (3) Network C1 and / or Network C2. In an optimized construct, Network D may be trained directly (without Networks A, B, C1, and C2). Thus, training for the different options presented herein includes (1) training Networks A, B, and C1; (2) training Networks A, B, and C2; (3) training Networks A, B, C1, and C2 (then selecting the best of C1 and C2 to perform inference using Networks A and B); and (4) training only Network D.

[0092] Regarding the training of networks C1 and / or C2, networks C1 and C2 are configured to have the same functionality with identical outputs. The difference between them is the input. Network C1 uses the outputs of networks A and B as input, while network C2 uses the intermediate latent features of networks A and B as input. The embodiments described herein train both networks C1 and C2, then compare their performance, and select the preferred model for inference.

[0093] In some embodiments, the pre-trained DefectGPT encoder is pre-trained once, thereby enabling the determination of different types of information from one or more different types of inputs without additional training. For example, one of the most important advantages of the embodiments described herein is the ability to pre-train once and apply the embodiments to different semiconductor steps, physical IC designs, processes for defect detection, or defect synthesis (as opposed to previously requiring different models to be trained for each step, physical IC design, or process). Furthermore, the pre-trained DefectGPT encoder in each of networks A, B, C1, C2, and D used for the defect synthesis and defect detection embodiments can be identical. The decoder architecture configured in each of the networks used for defect synthesis can also be identical, and it can be pre-trained or randomly initialized.

[0094] In another embodiment, the pre-trained DefectGPT encoder is configured to determine information by synthesizing DOIs for more than one optical mode. For example, another advantage of the embodiments described herein is the ability to synthesize DOIs in a multi-optical-mode context, taking into account process and tooling conditions.

[0095] In another embodiment, the pre-trained DefectGPT encoder is configured to determine information by synthesizing a DOI distribution in a Bayesian perspective. For example, a key advantage of the embodiments described herein is the ability to synthesize a DOI distribution in a Bayesian perspective. Here, a Bayesian perspective refers to the fact that the simulated images from networks C1 and C2 can be an imaging distribution rather than a 2D matrix with a value at each location. The imaging distribution can be approximated by, for example, a multivariate Gaussian distribution or other multivariate 2D distributions.

[0096] In an additional embodiment, the pre-trained DefectGPT encoder is configured to determine information by synthesizing process noise distributions for more than one optical mode. For example, another advantage of the embodiments described herein is the ability to synthesize process noise distributions in a multi-optical-mode context, taking into account process and tooling conditions. The process noise distribution can be generated by the Bayesian distribution described above, as part of the covariance or variation matrix. The desired model (e.g., networks C1 and C2) predicts the noise properties using Bayesian statistics.

[0097] In one embodiment, one or more components include a decoder configured to generate synthetic defect information from information determined for a sample by a pre-trained DefectGPT encoder, and a computer subsystem configured to train a defect detection model using the synthetic defect information. For example, defect images generated by the embodiments described herein can be used to train AI-based detection models for optical inspection, such as those described in U.S. Patent No. 11,551,348, published January 10, 2023, and U.S. Patent Application Publication No. 2024 / 0013365, published January 11, 2024, by Zhang et al., which are incorporated herein by reference as if fully described herein. The embodiments described herein may be further configured as described in those references. In another embodiment, the pre-trained DefectGPT encoder and the defect detection model learn jointly, which may be performed in any suitable manner known in the art. In yet another embodiment, the pre-trained DefectGPT encoder and the defect detection model are learned through supervised fine-tuning. For example, joint learning can be performed by directly training a pre-trained DefectGPT encoder and defect detection model or by learning through supervised fine-tuning (SFT), which can be done in any suitable manner known in the relevant field.

[0098] In another embodiment, the pre-trained DefectGPT encoder is configured for learning guided by the defect detection model via reinforcement learning. In other words, the DefectGPT model learning can be guided by the defect detection model via reinforcement learning, which can be performed in any suitable manner known in the relevant field.

[0099] In an additional embodiment, a pre-trained DefectGPT encoder is configured to determine information for single-mode or multi-mode optical conditions. Training an AI-based defect model for optical inspection using defect images generated as described herein offers several benefits, such as improving the sensitivity of defect detection models in single-mode and / or multi-mode optical inspection. Another benefit of this training is its ability to improve the detection of out-of-distribution (OOD) defects. Here, an out-of-distribution (OOD) defect refers to a potential defect that an application or customer expects to be present on a wafer but is difficult to detect or identify by other methods. Because these types of defects are generally not available for downstream tasks, such as defect detection recipe building, they are not part of the candidate or dataset during the build process, which is why they are OOD. By using the embodiments described herein, the proposed method can simulate these “expected” defects and use them for downstream tasks, such as training a defect detector. The detector can then be used to increase the chance of detecting these OOD defects in real-world experiments. An additional benefit is that this training can improve the generalization of the AI ​​detection model, for example, across different wafers and / or different processes.

[0100] In another embodiment, the pre-trained DefectGPT encoder is configured to determine information by simultaneously generating pattern and defect information of the sample. For example, as further described herein, (1) a combination of networks A, B, and C1, (2) a combination of networks A, B, and C2, and (3) network D can simultaneously generate pattern and defect information of the sample.

[0101] In some embodiments, the pre-trained DefectGPT encoder is configured to determine information in response to process variations, process conditions, and tooling conditions. For example, embodiments may consider any one or more of process variations, process conditions, and tooling conditions.

[0102] In another embodiment, the pre-trained DefectGPT encoder is configured to determine information by determining the defect distribution. In yet another embodiment, the pre-trained DefectGPT encoder is configured to determine information by determining the noise distribution. For example, in a Bayesian perspective or distribution, the defect and noise distributions can be determined as further described herein.

[0103] In some embodiments, the computer subsystem is configured to pre-train an initial DefectGPT encoder without any real DOI instances (or with a finite number of real DOI instances), thereby generating a pre-trained DefectGPT encoder, and the computer subsystem is configured to detect defects on a sample, generate synthetic defect information, or a combination thereof, based on information determined for the sample by the pre-trained DefectGPT encoder. For example, the embodiments described herein can operate with zero or a finite number of real DOIs.

[0104] 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 synthetic defect information, images of generated defect signals, simulated images of samples with and / or without defects, etc., for training ML or DL ​​models or networks, such as those described herein. In one such example, the computer subsystem may store this information and / or images in a training data structure or file. Information, images, etc., may be stored in any suitable manner in any computer-readable storage medium described herein.

[0105] In another example, the computer subsystem may be configured to store DOI query embeddings, synthetic defect information, generated defect signal images, simulated images of samples with and / or without defects, ML or DL ​​models or networks trained using any such information, and the like, for use during a process performed on a sample (such as those described herein). The computer subsystem may be configured to store this information and / or model or network in a recipe or by generating a recipe for a process that uses the information and / or model or network. As used herein, the term "recipe" 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 recipe may include information on how to perform the process, which can then be used to generate instructions for performing the process. The computer subsystem may also store any such information and / or model or network that can be used to identify, access, and / or use (e.g., file names and their storage locations). The stored information about the model or network may also include the model or network's code, instructions, algorithms, etc. Therefore, information, models, or networks, etc., may be stored in any suitable manner on any computer-readable storage medium described herein.

[0106] Information, models, or networks 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 information has been stored, it may be accessed in the storage medium 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, embodiments described herein may generate test recipes, as described above. The test recipes may then be stored and used by a system or method (or another system or method) to test samples or other samples to generate information about the samples or other samples (e.g., defect information). A computer subsystem may also be configured to detect defects on samples, as described herein, and information generated by the computer subsystem for detected defects may be stored and used, as further described herein.

[0107] The results and information generated by performing inspections on 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 manufacturing processes or steps that have been performed on or will be performed on the sample being inspected or another sample. For example, a computer subsystem can be configured to determine one or more modifications to the process that has been performed on or will be performed on a sample as described herein, based on detected defects. The modification to the process may include any suitable changes to one or more parameters of the process. The computer subsystem preferably determines the modification such that defects can be reduced or prevented on other samples to which a modified process is performed, defects on the sample can be corrected or eliminated in another process performed on the sample, defects can be compensated for in another process performed on the sample, and so on. The computer subsystem can determine such modifications in any suitable manner known in the art.

[0108] The changes can then be sent to a semiconductor manufacturing system (not shown) or storage medium (not shown) accessible to the computer subsystem and 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, for example, via one or more common elements (e.g., 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.

[0109] Therefore, as described herein, the embodiments can be used to establish new testing procedures or formulations. The embodiments can also be used to modify existing testing procedures or formulations, whether they are testing procedures or formulations for a sample or testing procedures or formulations created for one sample and adapted to another.

[0110] The embodiments described herein are not limited to verifying the creation or modification of recipes or processes. For example, the embodiments described herein can also be used to create or modify recipes or processes for metrology, defect re-detection, etc., in a similar manner. Specifically, the pre-trained DefectGPT encoder described herein may be trained depending on the process being created or modified (e.g., to generate simulated output that mimics the actual output to be produced by said process). Then, depending on the process or recipe being created or modified, the simulated output can be used to create a recipe for said process, whether it is synthetic information and / or simulated images stored in the process or used to train a DL or ML model or network used in the process. Such output processing methods may include, for example, defect re-detection methods for re-detecting defects in the output generated by a defect re-detection system.

[0111] Similarly, the embodiments described herein can be used not only to select output processing parameters and methods, but also to select output acquisition parameters or modes, used, for example, in inspection systems, metrology systems, or defect re-inspection systems to detect light, electrons, ions, etc., from a sample. This selection of output acquisition parameters may involve generating simulated images using a pre-trained DefectGPT encoder for different output acquisition parameters or modes, which can be performed as further described herein. The generated simulated images can then be compared and evaluated to select which modes or parameters are best for any given process. Therefore, the embodiments described herein can be used not only to establish or modify inspection processes, but also to establish or modify any type of quality control process performed on the samples described herein, and any parameters of that process.

[0112] Each of the embodiments described above can be combined to form a single embodiment. In other words, unless otherwise mentioned herein, no embodiment is mutually exclusive with any other embodiment.

[0113] Another embodiment relates to a computer-implemented method for determining information about a sample. The method includes inputting one or more sample-specific inputs into a pre-trained DefectGPT encoder, the pre-trained DefectGPT encoder being configured to determine information about the sample based on the one or more inputs. The inputs are executed by a computer subsystem. One or more components are executed by the computer subsystem, and one or more components include the pre-trained DefectGPT encoder.

[0114] Each of the steps in the method may be performed as further described herein. The method may also include any other steps that can be performed by the inspection system and / or computer system described herein. Furthermore, the method described above may be performed by any of the system embodiments described herein.

[0115] Additional embodiments relate to a non-transitory computer-readable medium storing program instructions executable on a computer system to perform a computer-implemented method for determining information about a sample. Figure 11 An example of this embodiment is shown below. Specifically, as... Figure 11 As shown, the non-transitory computer-readable medium 1100 contains program instructions 1102 that can be executed on a computer system 1104. A computer-implemented method may include any step of any method described herein.

[0116] Program instructions 1102 for implementing a method (such as the method described herein) may be stored on a computer-readable medium 1100. The computer-readable medium may be a storage medium, such as a disk or optical disc, magnetic tape, or any other suitable non-transitory computer-readable medium known in the art.

[0117] 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), Python, Tensorflow, or other technologies or methods as needed.

[0118] Computer system 1104 may be configured according to any of the embodiments described herein.

[0119] In view of 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 should be interpreted only as illustrative and for the purpose of teaching those skilled in the art the general manner of practicing the invention. It should be understood that the forms of the invention shown and described herein should be considered as the presently preferred embodiments. As those skilled in the art will understand after benefiting from this description of the invention, elements and materials may be substituted for those described herein, parts and processes may be reversed, and certain properties 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 following claims.

Claims

1. A system configured to determine information about a sample, comprising: Computer subsystem; and One or more components, executed by the computer subsystem, wherein the one or more components include a pre-trained defect-generating pre-trained transformer (DefectGPT) encoder, the pre-trained DefectGPT encoder being configured to determine information about the sample based on one or more sample-specific inputs, and wherein the computer subsystem is configured to input the one or more inputs into the pre-trained DefectGPT encoder.

2. The system of claim 1, wherein the computer subsystem is further configured to detect defects on the sample based on the information determined by the pre-trained DefectGPT encoder for the sample.

3. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder, thereby generating the pre-trained DefectGPT encoder, and wherein the pre-training is performed using a training dataset comprising images specific to the sample and images unrelated to the information used to determine the sample based on the one or more inputs.

4. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder, thereby generating the pre-trained DefectGPT encoder, and wherein the pre-training is performed using a training dataset that includes only images specific to the sample.

5. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder, thereby generating the pre-trained DefectGPT encoder, and wherein the initial DefectGPT encoder is configured to encode an input image into a visual lexical embedding.

6. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder in a self-supervised manner, thereby generating the pre-trained DefectGPT encoder.

7. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the defect of interest (DOI) query embedding based on one or more input DOI cue images.

8. The system of claim 7, wherein one or more inputs include a target image for defect detection, wherein determining the information includes determining a visual lexical embedding of the target image, and wherein the computer subsystem is further configured to determine a similarity metric between the DOI query embedding and the visual lexical embedding, and to determine whether a defect exists in the target image based on the determined similarity metric.

9. The system of claim 1, wherein one or more components further include a decoder configured to generate synthetic defect information from the information determined by the pre-trained DefectGPT encoder for the sample.

10. The system of claim 1, wherein the computer subsystem is further configured to generate fused DOI features from DOI location hints and DOI descriptors, wherein one or more inputs include the fused DOI features, wherein the determined information includes DOI embeddings, and wherein one or more components further include a decoder configured to generate a defect signal image from the DOI embeddings.

11. The system of claim 1, wherein the computer subsystem is further configured to generate a fused pattern feature from a design image, optical pattern information, and process parameter information of the sample, wherein one or more inputs include the fused pattern feature, wherein the determined information includes a pattern embedding, and wherein the one or more components further include a decoder configured to generate a defect-free simulated image of the sample from the pattern embedding.

12. The system of claim 1, wherein the computer subsystem is further configured to generate a fusion input from a defect signal image of the sample and a non-defective simulated image of the sample, wherein one or more inputs include the fusion input, wherein the determined information includes a simulated embedding, and wherein the one or more components further include a decoder configured to generate a simulated image of the sample having defects from the simulated embedding.

13. The system of claim 1, wherein the determined information includes DOI embedding, pattern embedding, and combined embedding generated from the DOI embedding and the pattern embedding, and wherein one or more components further include a decoder configured to generate a simulated image of defects on the sample from the combined embedding.

14. The system of claim 1, wherein the one or more inputs include a design image of the sample, and wherein the one or more components further include a first diffusion process model configured to generate a design embedding from the information determined by the pre-trained DefectGPT encoder for the sample.

15. The system of claim 14, wherein the one or more components further include a condition encoder configured to determine additional information about the sample from DOI location cues, DOI descriptors, optical mode information, and process parameter information, and wherein the one or more components further include a second diffusion process model configured to generate conditionally embedded data from the additional information.

16. The system of claim 15, wherein one or more components further include a decoder configured to generate a simulated image of the sample with defects based on the design embedding and the conditional embedding.

17. The system of claim 1, wherein the pre-trained DefectGPT encoder is configured as a Vision Transformer (ViT) or a Swing Transformer.

18. The system of claim 1, wherein one or more components further include a decoder configured to determine additional information about the sample from the information determined for the sample, and wherein the decoder is configured as a vision transformer (ViT) or a Swing transformer.

19. The system of claim 1, wherein the pre-trained DefectGPT encoder is pre-trained once, thereby enabling the determination of different kinds of information from different kinds of the one or more inputs without additional training.

20. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by synthesizing a DOI for more than one optical mode.

21. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by synthesizing a DOI distribution in a Bayesian perspective.

22. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by means of noise distribution for synthesis processes of more than one optical mode.

23. The system of claim 1, wherein one or more components further include a decoder configured to generate synthetic defect information from the information determined by the pre-trained DefectGPT encoder for the sample, and wherein the computer subsystem is further configured to use the synthetic defect information to train a defect detection model.

24. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by simultaneously and jointly generating pattern and defect information of the sample.

25. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information for single-mode or multi-mode optical conditions.

26. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information in response to process variations, process conditions, and tooling conditions.

27. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by determining the defect distribution.

28. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured to determine the information by determining the noise distribution.

29. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder without having any real DOI instances, thereby generating the pre-trained DefectGPT encoder, and wherein the computer subsystem is further configured to detect defects on the sample, generate synthetic defect information, or a combination thereof, based on the information determined by the pre-trained DefectGPT encoder for the sample.

30. The system of claim 1, wherein the computer subsystem is further configured to pre-train an initial DefectGPT encoder using a finite number of real DOI instances, thereby generating the pre-trained DefectGPT encoder, and wherein the computer subsystem is further configured to detect defects on the sample, generate synthetic defect information, or a combination thereof, based on the information determined by the pre-trained DefectGPT encoder for the sample.

31. The system of claim 1, wherein the pre-trained DefectGPT encoder is jointly learned with the defect detection model.

32. The system according to claim 1, wherein the pre-trained DefectGPT encoder and defect detection model are learned through supervised fine-tuning.

33. The system of claim 1, wherein the pre-trained DefectGPT encoder is further configured for learning guided by the defect detection model via reinforcement learning.

34. A non-transitory computer-readable medium storing program instructions executable on a computer system to perform a computer-implemented method for determining information about a sample, wherein the computer-implemented method includes: One or more sample-specific inputs are fed into a pre-trained DefectGPT encoder, which is configured to determine information about the sample based on the one or more inputs.

35. A computer-implemented method for determining information about a sample, comprising: One or more sample-specific inputs are fed into a pre-trained DefectGPT encoder, which is configured to determine information about the sample based on the one or more inputs, wherein the inputs are executed by a computer subsystem, wherein one or more components are executed by the computer subsystem, and wherein the one or more components include the pre-trained DefectGPT encoder.

Images