Mode selection and defect detection training

Abstract

A system can be configured for joint defect discovery and optical mode selection. Defects are detected during a defect discovery step. Discovered defects are accumulated into a mode selection dataset. Mode selection is performed using the mode selection dataset to determine a mode combination. The mode combination can then be used to train the defect detection model. Additional defects can then be detected by the defect detection model. The additional defects can then be provided to the mode selection dataset to further perform mode selection and train the defect detection model. One or more run-time modes can then be determined. The system can be configured for image pixel level mode selection and defect detection.

Description

[0001] Cross-citation of related applications

[0002] This application asserts the rights of U.S. Provisional Application No. 63 / 027,975, filed May 21, 2020, pursuant to 35 USC §119(e), the entire contents of which are incorporated herein by reference. Technical Field

[0003] This invention generally relates to semiconductor inspection, and more specifically to the classification of defects detected by semiconductor inspection. Background Technology

[0004] Semiconductor manufacturing environments are typically highly controlled to prevent wafer contamination by foreign matter that could interfere with the manufacturing process or degrade the performance of the fabricated devices. Inspection systems are commonly used to locate defects, such as (but not limited to) foreign particles, on the substrate for screening and avoidance measures. The sensitivity of defect inspection can vary based on factors such as (but not limited to) defect type, measurement parameters, or defect detection models. Therefore, identifying suitable measurement parameters and defect detection models can be challenging.

[0005] Therefore, it would be advantageous to provide a system and method that addresses the shortcomings described above. Summary of the Invention

[0006] A system according to one or more illustrative embodiments of the present disclosure is disclosed. In one illustrative embodiment, the system includes a controller. In another illustrative embodiment, the controller is communicatively coupled to an inspection subsystem. In another illustrative embodiment, the inspection subsystem is configured to image at least one sample when configured to have any of a plurality of candidate optical modes. In another illustrative embodiment, the controller includes one or more processors configured to execute program instructions that cause the one or more processors to jointly perform optical mode selection and defect detection training. In another illustrative embodiment, the processor receives defect data of at least one defect on at least a portion of the at least one sample. In another illustrative embodiment, the processor receives at least one image from the inspection subsystem and stores the at least one image in a dataset. In another illustrative embodiment, the processor selects one or more optical modes from the plurality of candidate modes by executing a mode selection model. In another illustrative embodiment, the processor trains a defect detection model using the image associated with the one or more optical modes selected by the mode selection model. In another illustrative embodiment, the processor is further configured to determine at least one runtime optical mode from the plurality of candidate optical modes.

[0007] A method according to one or more illustrative embodiments of the present disclosure is disclosed. The method may include performing optical pattern selection and defect detection training. In one illustrative embodiment, the method includes receiving defect data of at least one defect on at least a portion of at least one sample. In another illustrative embodiment, the method includes receiving at least one image from an inspection subsystem and storing the at least one image in a dataset. In another illustrative embodiment, the method includes selecting one or more optical patterns from a plurality of candidate optical patterns by performing a pattern selection model. In another illustrative embodiment, the method includes training a defect detection model using the images associated with the one or more optical patterns selected by the pattern selection model. In another illustrative embodiment, the method includes performing a defect inspection test.

[0008] A system according to one or more illustrative embodiments of the present disclosure is disclosed. In one illustrative embodiment, the system includes an inspection subsystem configured to image at least one sample when configured to have a plurality of candidate optical modes. In another illustrative embodiment, the system includes a controller communicatively coupled to the inspection subsystem. In another illustrative embodiment, the controller includes one or more processors configured to execute program instructions that cause the one or more processors to jointly perform optical mode selection and defect detection training. In another illustrative embodiment, the processor receives defect data of at least one defect on at least a portion of the at least one sample. In another illustrative embodiment, the processor receives at least one image from the inspection subsystem and stores the at least one image in a dataset. In another illustrative embodiment, the processor selects one or more optical modes from the plurality of candidate modes by executing a mode selection model. In another illustrative embodiment, the processor trains a defect detection model using the image associated with the one or more optical modes selected by the mode selection model. In another illustrative embodiment, the processor is further configured to determine at least one runtime optical mode from the plurality of candidate optical modes. Attached Figure Description

[0009] By referring to the accompanying drawings, those skilled in the art can better understand the many advantages of this disclosure, among which:

[0010] Figure 1A This is a conceptual diagram of an image-based optical inspection system according to one or more embodiments of the present disclosure;

[0011] Figure 1B This is a conceptual diagram of an image-based optical inspection system according to one or more embodiments of the present disclosure;

[0012] Figure 1CThis is a simplified schematic diagram of an image-based optical inspection system according to one or more embodiments of the present disclosure;

[0013] Figures 2A to 2B A flowchart depicting a method for joint defect detection and optical mode selection according to one or more embodiments of the present disclosure;

[0014] Figure 3 Describe a pattern selection model according to one or more embodiments of the present disclosure;

[0015] Figure 4 Describing a mode selection model according to one or more embodiments of the present disclosure; and

[0016] Figure 5 A training defect detection model is described according to one or more embodiments of the present disclosure. Detailed Implementation

[0017] The disclosed subject matter will now be described in detail with reference to the accompanying drawings.

[0018] Defect inspection of samples can be performed using a wide range of tools, including optical inspection tools and electron beam tools. Typically, optical inspection tools offer higher throughput, while electron beam tools offer higher resolution. Optical inspection tools may include one or more adjustable optical characteristics. A combination of optical characteristics of an inspection tool can be referred to as a recipe or optical mode; such characteristics include (but are not limited to) wavelength, focal length, aperture, or bandwidth. Optical inspection tools may contain hundreds or thousands of optical modes, some of which produce images more or less suitable for defect detection.

[0019] Defect inspection can be performed to detect defects in an image by applying a defect detection model to an image of a sample acquired through an optical inspection tool. Defect inspection may involve comparing the image with one or more reference images (e.g., a reference image acquired by another tool, a reference image acquired by the optical inspection tool configured with different optical modes, a reference image acquired from another die on the sample, reference images acquired from one or more reference dies, a combination of one or more of the aforementioned reference images, etc.). Some defect detection models are more or less suitable for detecting defects in images acquired in a given optical mode. Therefore, it is desirable to have optical modes and defect detection models that can cooperatively detect defects on samples during production steps or other runtime environments. In a general sense, any number of optical modes can be used to inspect samples during runtime. However, using different optical modes for inspection during runtime typically negatively impacts production throughput. Therefore, the number of optical modes used during runtime is typically limited (e.g., limited to 1 to 3 modes). However, it is difficult to determine a suitable combination of optical modes and defect detection models a priori.

[0020] Embodiments of this disclosure relate to joint defect detection and optical mode selection. Data sufficiency for optical mode selection during defect training can be addressed by performing joint defect detection and optical mode selection. Furthermore, it is possible to adjust the optical mode selection for a desired number of candidate optical modes.

[0021] Joint defect detection and optical pattern selection may involve identifying one or more defects (e.g., defects of interest (DOI)) on a sample. Furthermore, an inspection subsystem may generate one or more images associated with the defects, wherein this inspection subsystem is configured to have optical patterns. The images and associated defects can be accumulated into a pattern selection dataset. The pattern selection dataset can be used with a pattern selection algorithm to determine one or more optical patterns. The one or more optical patterns can then be used to train a defect detection model. Optionally, one or more optical patterns can be used with an optical inspection tool to perform performance evaluation at one or more locations at the die, die row, or wafer level. Defects discovered during performance evaluation can be further accumulated in the pattern selection dataset for subsequent iterative iterations.

[0022] In this embodiment, defect detection and optical mode selection can be performed iteratively. By iteratively performing defect detection and optical mode selection, a defect detection model can be trained, and at least one optical mode can be selected. During runtime, the inspection subsystem can acquire at least one image of a sample using at least one optical mode, and the defect detection model can detect one or more defects in at least one image. Furthermore, defect detection using the defect detection model can be performed at the desired sensitivity and stability level.

[0023] A system and method for generating test formulations are described in U.S. Patent No. 7,877,722 to Brian Duffy, the entire contents of which are incorporated herein by reference.

[0024] The classification of multimode defects in semiconductor inspection is described in U.S. Patent Application No. 16 / 272,528, published by Vaibhav Gaind as US 2020 / 0025689, the entire contents of which are incorporated herein by reference.

[0025] General reference Figures 1A to 5 The present disclosure discloses a system 100 and a method 200 according to one or more embodiments thereof.

[0026] Figure 1A This is a conceptual diagram illustrating an image-based optical inspection system 100 according to one or more embodiments of the present disclosure.

[0027] System 100 may include (but is not limited to) an inspection subsystem 102. The inspection subsystem 102 may be configured in multiple candidate optical modes. For example, in the optical modes of multiple candidate optical modes, individual stacked target elements may be resolved within the illumination spot on the sample (e.g., as part of a bright-field image, dark-field image, phase-contrast image, or the like). The optical mode may include a combination of optical characteristics of the inspection subsystem. The optical mode of the inspection subsystem 102 may include (but is not limited to) the illumination wavelength, the detection wavelength of radiation emitted from the sample, the size of the illumination spot on the sample, the angle of incident illumination, the polarization of the incident illumination, the focal length, the position of the incident illumination beam on the stacked target, or the transmission distribution in the focusing aperture. In this respect, the optical mode may be associated with the formulation of the inspection subsystem 102. For example, the optical mode includes aperture, wavelength, and polarization. If ten different apertures, ten different wavelengths, and three polarizations are considered, then there are three hundred possible optical modes (e.g., candidate optical modes).

[0028] In one embodiment, controller 101 is communicatively coupled to inspection subsystem 102. In another embodiment, controller 101 is configured to generate and provide one or more control signals configured to perform one or more adjustments (e.g., adjusting the wavelength of inspection subsystem 102) on one or more portions of inspection subsystem 102.

[0029] In an embodiment, controller 101 may also be configured to receive image data from inspection subsystem 102. Controller 101 may also be configured to perform any of the various steps further discussed herein. For example, the controller may include a mode selection model 103 and a defect detection model 104. Through mode selection model 103 and defect detection model 104, controller 101 may jointly perform optical mode selection and defect detection. Joint optical mode selection and defect detection can address the lack of data associated with optical mode selection (e.g., when not all optical modes have associated images). Furthermore, defect detection model 104 may be sufficiently tuned to be trained using deep learning techniques and can achieve sufficient detection when few defects are available. By jointly performing optical mode selection and defect detection, one or more runtime optical modes and a trained defect detection model can be determined, which can then be provided to inspection subsystem 102 during runtime.

[0030] Figure 1B This is a conceptual diagram illustrating one or more embodiments of a system according to the present disclosure.

[0031] The controller 101 may include one or more processors 105, memory 107, and may include or be coupled to the user interface 110. The one or more processors 105 of the controller 101 may perform any of the various processing steps described throughout this disclosure, such as (but not limited to) joint execution mode selection and defect detection.

[0032] One or more processors 105 of controller 101 may comprise any processor or processing element known in the art. For the purposes of this disclosure, the terms “processor” or “processing element” may be broadly defined to encompass any device having one or more processing or logic elements (e.g., one or more microprocessor devices, one or more application-specific integrated circuit (ASIC) devices, one or more field-programmable gate arrays (FPGAs), or one or more digital signal processors (DSPs)). In this sense, one or more processors 105 may comprise any device configured to execute algorithms and / or instructions (e.g., program instructions stored in memory). In embodiments, one or more processors 105 may embody a desktop computer, host computer system, workstation, graphics computer, parallel processor, networked computer, or any other computer system configured to execute a program configured to operate with or in conjunction with operating system 100, as described throughout this disclosure. Furthermore, different subsystems of system 100 may include processors or logic elements adapted to perform at least a portion of the steps described in this disclosure. Therefore, the foregoing description should not be construed as limiting the embodiments of this disclosure but is merely illustrative. Furthermore, the steps described throughout this disclosure can be performed by a single controller or alternatively by multiple controllers. Additionally, controller 101 may comprise one or more controllers housed in a common enclosure or within multiple enclosures. In this manner, any controller or combination of controllers can be individually packaged as a module suitable for integration into system 100. Furthermore, controller 101 can analyze data received from the verification subsystem 102 and feed the data to additional components within or outside system 100.

[0033] Memory media 107 may comprise any storage medium known in the art suitable for storing program instructions executable by one or more associated processors 105. For example, memory media 107 may comprise a non-transitory memory medium. As another example, memory media 107 may comprise (but is not limited to) read-only memory (ROM), random access memory (RAM), magnetic or optical storage devices (e.g., magnetic disks), magnetic tape, solid-state drives, and the like. It should be further noted that memory media 107 may be housed within a common controller housing with one or more processors 105. In embodiments, memory media 107 may be remotely located relative to the physical location of one or more processors 105 and controller 101. For example, one or more processors 105 of controller 101 may access remote memory (e.g., a server) accessible via a network (e.g., the Internet, an intranet, and the like).

[0034] In one embodiment, the user interface is communicatively coupled to the controller 101. In another embodiment, the user interface 110 may include (but is not limited to) one or more desktop computers, laptop computers, tablet computers, and the like. In another embodiment, the user interface 110 includes a display for displaying data from the system 100 to a user. The display of the user interface 110 may include any display known in the art. For example, the display may include (but is not limited to) a liquid crystal display (LCD), an organic light-emitting diode (OLED) based display, or a CRT display. Those skilled in the art will recognize that any display device capable of being integrated with the user interface 110 is suitable for implementation in this disclosure. In another embodiment, a user may input selections and / or commands in response to data displayed to the user via a user input device of the user interface 110.

[0035] Figure 1C A simplified schematic diagram illustrating an image-based optical imaging system 100 according to one or more embodiments of the present disclosure.

[0036] The inspection subsystem is further described in U.S. Patent Application No. 16 / 572,971, published by Ramaprasad Kulkarni as US 2020 / 0143528, the entire contents of which are incorporated herein by reference.

[0037] The inspection subsystem 102 may include any inspection subsystem known in the art, including (but not limited to) imaging-based optical inspection subsystems. For the purposes of this disclosure, the term "metrology tool" may be used interchangeably with "inspection subsystem". In this sense, the metrology tool and controller 101 may form system 100. Furthermore, the inspection subsystem 102 may include any type of optical metrology tool known in the art suitable for generating metrological data.

[0038] In one embodiment, the inspection subsystem 102 includes an illumination source 106 configured to generate an illumination beam 108. The illumination beam 108 may contain one or more selected wavelengths of light, including (but not limited to) vacuum ultraviolet (VUV), deep ultraviolet (DUV), ultraviolet (UV) radiation, visible radiation, or infrared (IR) radiation. The illumination source 106 may further generate an illumination beam 108 containing any range of selected wavelengths. In another embodiment, the illumination source 106 may include a spectrally tunable illumination source to generate an illumination beam 108 having a tunable spectrum.

[0039] The illumination source 106 may further generate an illumination beam 108 having any time profile. For example, the illumination source 106 may generate a continuous illumination beam 108, a pulsed illumination beam 108, or a modulated illumination beam 108. In addition, the illumination beam 108 may be delivered from the illumination source 106 via free-space propagation or guiding light (e.g., optical fiber, optical guide, or the like).

[0040] In another embodiment, illumination source 106 guides illumination beam 108 to sample 111 via illumination path 109. Illumination path 109 may include illumination optics 114 adapted to modify and / or adjust illumination beam 108. For example, one or more illumination optics 114 may include (but are not limited to) one or more lenses, one or more polarizers, one or more filters, one or more diffusers, one or more homogenizers, one or more apodizers, one or more beam shapers, or one or more shutters (e.g., mechanical shutters, electro-optic shutters, acousto-optic shutters, or the like). As another example, one or more illumination optics 114 may include aperture stops for controlling the illumination angle on sample 111 and / or field stops for controlling the spatial range of illumination on sample 111. In another embodiment, system 100 includes beam splitter 120. In another embodiment, system 100 includes objective lens 116 for focusing illumination beam 108 onto sample 111.

[0041] In another embodiment, sample 111 is placed on sample stage 118. Sample stage 118 may include any means suitable for positioning sample 111 within system 100. For example, sample stage 118 may include any combination of a linear translation stage, a rotation stage, a tilting / tilting stage, or the like.

[0042] In another embodiment, detector 112 is configured to capture radiation emitted from sample 111 via collection path 122. For example, collection path 122 may include (but does not necessarily include) a light-collecting lens (e.g., such as...). Figure 1B(Objective 116 as described in the diagram). In this respect, detector 112 may receive radiation reflected or scattered from sample 111 (e.g., via specular reflection, diffuse reflection and the like) or generated through sample 111 (e.g., luminescence or the like associated with absorption of illumination beam 108).

[0043] The collection path 122 may further include any number of collection optics 124 for guiding and / or modifying the illumination collected by the objective lens 116, including (but not limited to) one or more collection path lenses, one or more filters, one or more polarizers, or one or more beam stops. Furthermore, the collection path 122 may include a field stop for controlling the spatial extent of the sample imaged onto the detector 112, or an aperture stop for controlling the angular extent of the illumination from the sample used to generate an image on the detector 112. In another embodiment, the collection path 122 includes an aperture stop positioned in a plane conjugate to the back focal plane of the optics of the objective lens 116 to provide telecentric imaging of the sample. In an embodiment, the inspection subsystem 102 includes a beam splitter 120 oriented such that the objective lens 116 can simultaneously guide the illumination beam 108 to the sample 111 and collect radiation emitted from the sample 111.

[0044] Detector 112 may comprise any type of optical detector known in the art suitable for measuring illumination received from sample 111. For example, detector 112 may comprise (but is not limited to) a charge-coupled device (CCD) detector, a time-delay integration (TDI) detector, a photomultiplier tube (PMT), an avalanche photodiode (APD), a complementary metal-oxide-semiconductor (CMOS) sensor, or the like. In another embodiment, detector 112 may comprise a spectral detector suitable for identifying the wavelength of light emitted from sample 111.

[0045] In this embodiment, controller 101 is communicatively coupled to detector 112. Therefore, controller 101 can receive dataset 126 from detector 112. Dataset 126 may contain one or more images 128 acquired according to the optical mode of inspection subsystem 102. Dataset 126 can be generated by receiving multiple images from at least a portion of sample 111 and / or from multiple samples (e.g., bare pieces or rows of bare pieces). For example, multiple images may be acquired during a defect inspection test. This defect inspection test can be performed when inspection subsystem 102 is configured to have an optical mode (e.g., such that the optical mode can be associated with images 128 in dataset 126).

[0046] In an embodiment, the dataset may further include defect data associated with one or more images 128. This defect data may include one or more defects detected on the samples (e.g., pixel-level information from samples from an electron beam inspection (EBI) tool or other tool). In an embodiment, the controller 101 may receive at least a portion of the dataset 126 from another inspection tool (e.g., receiving defect data from an electron beam inspection (EBI) tool or other tool). This defect data may then be associated with one or more images 128, allowing the defect data to be used to train a defect detection model by utilizing the location of defects on the images.

[0047] In this embodiment, dataset 126 may be stored in and received from memory 107. As will be understood, any number of optical patterns, images, and defect data may be accumulated in dataset 126 and subsequently accumulated in memory 107.

[0048] In an embodiment, when training using a combination of one or more optical modes, the defect detection model 104 can be determined to have the desired quality (e.g., sensitivity or stability). The combination of one or more optical modes can be determined by a mode selection model 103. The mode selection model 103 can determine the combination of one or more optical modes by selecting a portion of the dataset 126 used to train the defect detection model 104. A portion of the dataset 126 used to train the defect detection model 104 may contain images acquired from regions of previously imaged samples (e.g., by different optical modes) or from regions of samples that have not been previously imaged. When images are acquired from regions of samples previously used to train the defect detection model 104, the sensitivity of the defect detection model 104 can be evaluated. When images are acquired from regions of samples that have not been previously used, the stability of the defect detection model 104 can be evaluated. As will be understood, the defect detection model 104 should be trained for both stability and sensitivity.

[0049] In one embodiment, controller 101 may determine that dataset 126 does not contain a sufficient number of defect data. Then, controller 101 may receive additional defect data from coupling with electron beam inspection (EBI) or other such tools.

[0050] In an embodiment, controller 101 may also determine that dataset 126 does not contain a sufficient number of images (e.g., data for the desired optical pattern), and the controller may provide the desired optical pattern (e.g., a recipe) to inspection subsystem 102. Inspection subsystem 102 may then be configured according to the desired optical pattern and one or more images may be acquired according to the desired optical pattern. Similarly, controller 101 may determine a portion of sample 111 imaged above it or different samples imaged above it and provide this information to inspection subsystem 102. This determination may be based at least in part on defect data regarding defects detected by additional tools (e.g., EBI tools) and may optionally include region of interest analysis. The ability to provide additional desired optical patterns or portions of samples during training of pattern selection model 103 and defect detection model 104 ensures data sufficiency.

[0051] Figures 2A to 2B A flowchart illustrating the steps performed in a method 200 for performing joint defect detection and optical mode selection according to one or more embodiments of the present disclosure is provided. The embodiments and implementation techniques previously described in the context of system 100 should be interpreted as extending to method 200. However, it should be further understood that method 200 is not limited to system 100.

[0052] Using method 200, multiple candidate optical modes (e.g., hundreds or more modes from an inspection subsystem or multiple physical tools) can be narrowed down to one or more runtime modes. The runtime mode may include combinations of machine settings, such as (but not limited to) wavelength, polarization, focal length, bandwidth, aperture, transmission distribution in the illumination aperture, transmission distribution in the collecting aperture, and phase shift distribution in the collecting aperture, suitable for imaging one or more defects. Furthermore, through method 200, a defect detection model can be trained for the runtime mode.

[0053] In an embodiment, method 200 includes a defect detection step 202. Defect detection step 202 may include receiving defect data for at least one defect on at least a portion of at least one sample. Defects may be detected on one or more images generated from any source including (but not limited to) optical inspection tools (e.g., inspection subsystem 102), electron beam inspection tools, transmission electron microscopy (TEM) tools, or broadband plasma (BBP) inspection tools. Such images may be suitable for finding one or more defects on the sample, such as (but not limited to) voids, protrusions, or bridges. Defects may be detected on one or more images using conventional detection and image processing algorithms, pre-trained defect detection models (e.g., neural networks, deep learning models, etc.), or by a defect detection model trained during method 200. For example, defects associated with defect detection step 202 may be identified in one or more dies or dies of a sample inspected using a defect detection model (e.g., dies / dies rows inspected via defect inspection testing in step 210). Defect detection can be performed on one or more portions of a sample (e.g., sample 111) until a sufficient number of defects are found, such as (but not limited to) up to 30 defects or more. Furthermore, defect detection can be performed on multiple samples (e.g., dies, die rows, or wafers).

[0054] In an embodiment, method 200 includes a defect acquisition and accumulation step 204. The defect acquisition and accumulation step 204 may include receiving at least one image acquired by inspection subsystem 102 while configured according to a candidate optical mode, and storing this image in a dataset. The dataset may contain four-dimensional (4D) tensor data (or other arrays of this dimension). The tensor data may contain any suitable format (e.g., (N, W, H, C) format) corresponding to the number of images in a batch, the width of the images, the height of the images, and the number of channels in the images (e.g., channels for grayscale, three channels for red, green, and blue (RGB), etc.). The dataset may additionally be associated with defects detected in defect detection step 202 (e.g., based on the location of the defect on the sample). In this respect, at least one image may be from at least a portion of at least one sample for which at least one defect has been received.

[0055] In an embodiment, method 200 includes a step 206 of selecting one or more optical modes. The mode selection of one or more optical modes in step 206 may include executing a mode selection model (e.g., mode selection model 103) to select one or more optical modes from a plurality of candidate optical modes. One or more modes may be selected from a list of candidate modes. The list of candidate modes may contain only modes associated with defect data in dataset 126. Alternatively, the list of candidate modes may contain modes other than those included in the dataset (e.g., where the mode selection model is configured to select modes by interpolation). In cases where the list of candidate modes contains modes other than those included in the dataset, additional image collection may be required (e.g., via step 205).

[0056] In an embodiment, method 200 may optionally include step 205 of configuring a verification subsystem to have a selected optical mode and receiving an image acquired according to the selected optical mode. Step 205 may be performed after step 206 (e.g., after a mode selection model selects one or more modes). In this respect, the mode selection model may be configured to select one or more optical modes that do not have associated images in the dataset (e.g., due to insufficient data, the dataset does not contain images of all candidate modes; wherein this mode selection model may be configured to select modes that do not have images by interpolation). To address insufficient data, the verification subsystem may be configured to have a selected optical mode and acquire one or more images.

[0057] The pattern selection model used in step 206 may include at least one of the following: random channel drop vector, sparse vector, model agnostic meta-learning algorithm, forward selection algorithm, or backward selection algorithm. These pattern selection models are further discussed in this paper.

[0058] In an embodiment, method 200 includes a step 208 of training a defect detection model given a selected mode. Step 208 of training the defect detection model may include providing a dataset associated with the selected mode (e.g., images and defect data acquired according to the desired optical mode) to the defect detection model. The dataset may be used to train the defect detection model according to any suitable method (e.g., but not limited to, neural networks). The neural network may include at least one of deep generative models, convolutional neural networks (CNNs), generative adversarial networks (GANs), conditional generative adversarial networks (cGANs), variational autoencoders (VAEs), representation learning networks, or transformer models (e.g., bidirectional encoder representations from transformers (BERT); generative pre-trained transformers (GPT), etc.).

[0059] In an embodiment, method 200 includes a step 210 of performing a defect inspection test. The defect inspection test may involve inspecting one or more regions of one or more samples. In this regard, images of one or more samples may be acquired using inspection tools (e.g., inspection subsystem 102, electron beam inspection (EBI) tools, broadband plasma (BBP) inspection tools, or transmission electron microscopy (TEM) tools). Images may be acquired from regions of samples (e.g., sample 111) that have been previously imaged (e.g., to evaluate the stability of a defect detection model) or from regions of samples that have not been previously imaged (e.g., to evaluate the sensitivity of a defect detection model). In the case of acquiring a sample from a region of a previously unimaged sample, the region may be from the same sample or from a new sample (e.g., a die, a die row, or a wafer).

[0060] In an embodiment, method 200 may include step 211: performing inference on at least one of a die, die row, or wafer using a defect detection model to evaluate at least one of the stability or sensitivity of the defect detection model. The defect detection model can be evaluated by performing inference on one or more of the die, die row, or wafer level. The inference results can be compared with one or more known defect locations (e.g., determined via EBI tools, TEM tools, EBR tools, etc.). Based on the inference, the sensitivity and stability of the defect detection model can be evaluated. The defect detection model may require further training to detect defects in any given optical mode (e.g., candidate modes) using given stability and sensitivity. This training can be performed by iterating one or more steps of method 200 until stability or sensitivity criteria are met. Similarly, one or more steps of method 200 can be iterated until several iteration criteria are met.

[0061] Measurement formulation optimization based on spectral sensitivity and process variations is described in U.S. Patent No. 10,354,929 to Stilian Ivanov Pandev, the entire contents of which are incorporated herein by reference.

[0062] By combining defect detection and optical mode selection, a defect detection model can be configured to detect defects in sample images based on one or more optical modes used to train the model. Therefore, the defect detection model can be trained without considering the process of recording (POR). In this respect, the defect detection model (e.g., defect detection model 104) can detect additional pixels associated with one or more defects that are independent of feature quality.

[0063] In an embodiment, method 200 includes step 212 of determining at least one runtime optical mode. As will be understood, a minimum number of runtime optical modes is desired. In one instance, an optical mode is selected as the runtime optical mode. In another instance, up to three optical modes are selected as runtime optical modes. In this respect, the number of optical modes selected as runtime optical modes is not intended to be limited. In fact, any suitable number of optical modes can be selected as runtime optical modes. To determine the runtime optical modes, a ranking table can be generated. The ranking table may contain one or more indicators. A defect detection model can be applied to images in the dataset to determine defects in the images. Indicators can be generated, at least in part, based on the determined defects. Such indicators may be associated with the selection of runtime optical modes from the optical modes in the dataset. For example, indicators may include signal-to-noise ratio (SNR), receiver operating characteristic (ROC) curves, capture rate, interference rate, or computational cost. In this respect, the optical mode can be selected based on the intended application and / or indicators that are more relevant to the user. An image frame-based algorithmic selector is described in U.S. Patent Application No. 16 / 389,422, published by Bjorn Brauer as US 2020 / 0132610, the entire contents of which are incorporated herein by reference. Subsequently, at least one runtime optical mode and defect detection model can be provided to the inspection subsystem during runtime.

[0064] For reference Figure 3 and 4 Various embodiments of the pattern selection model 103 (or, for example, the pattern selection used in step 206) will be described in more detail according to one or more embodiments of this disclosure. The pattern selection model 103 may include one or more of the following: random channel drop vectors, sparse vectors, model-agnostic meta-learning algorithms, forward selection algorithms, or backward selection algorithms.

[0065] Figure 3 A pattern selection model 103 is depicted according to one or more embodiments of the present disclosure.

[0066] In an embodiment, the pattern selection model 103 includes a large model for random pattern inputs, learned based on dataset 126. The pattern set may include pattern 1, pattern 2, pattern 3, pattern 4, and at most pattern n (e.g., candidate patterns 1 to n). A subset of these patterns may have defect data (e.g., defect locations) and associated images 128 acquired by the verification subsystem 102 when configured with optical patterns. Such defect data and associated images 128 may be stored in dataset 126.

[0067] like Figure 3As depicted, mode selection model 103 may include random channel drop vector 302. Random channel drop vector 302 may have a length equal to the number of indices 304a to 304d. Such indices 304a to 304d may be associated with an optical mode having at least one image in dataset 126, or may be associated with all candidate optical modes (e.g., where one or more of the candidate modes do not have associated images in the dataset obtained by the verification subsystem when configured to have one or more candidate modes). Random channel drop indices 304a to 304d may contain mode selection values ​​randomly generated with zero or non-zero (e.g., 1) values. After generating random channel drop vector 302, random channel drop vector 302 may be applied to dataset 126 (e.g., grayscale channels of a 4D tensor, one or more RGB channels of a 4D tensor, etc.) to determine a subset 306 of dataset 126. In this respect, subset 306 may contain optical modes with associated non-zero random channel drop indices. Next, according to one or more embodiments, this subset 306 may be provided to the defect detection model 104 for training the defect detection model 104. Thus, the pattern selection model 103 can handle all combinations of patterns of interest (e.g., a list of candidate patterns).

[0068] Figure 4 A pattern selection model 103 is depicted according to one or more embodiments of the present disclosure.

[0069] In an embodiment, the pattern selection model 103 incorporates learning based on sparse constraints. For example, a sparse vector 402 may have index values ​​in a range (e.g., between 0 and 1). The length of the sparse vector 402 (e.g., the number of indices 404) may be equal to indices 404a to 404d. Such indices 404a to 404d may be associated with an optical pattern having at least one image in the dataset 126, or may be associated with all candidate optical patterns. The sparse vector 402 may be applied to the dataset 126 to determine a subset 406 of the dataset 126, wherein the subset 406 is used to train the defect detection model 104 by acting as a binary pattern selection indicator or as a weighted vector.

[0070] In this embodiment, the sparse vector 402 can be considered equivalent to a binary mode selection indicator. Initially (e.g., before finite constraints), it can be optimized using one or more sparse optimization techniques (e.g., but not limited to L1 optimization or L...). αOptimization (e.g., where 0 < α < 1) determines sparse vector 402. A subset 406 is then determined by comparing sparse vector indices 404a to 404d with a threshold (e.g., a threshold equal to 7). In this respect, any optical mode having an associated sparse vector index 404a to 404d greater than or equal to the threshold may be included in subset 406 (e.g., 4D tensor channels of the dataset). Subset 406 can then be provided to defect detection model 104 for training. It is understood that the threshold described herein is not intended to be restrictive. In this respect, any number (e.g., between 0 and 1) may be a suitable threshold. Furthermore, unless otherwise stated, the threshold should not be limited to values ​​greater than or equal to. In this respect, the threshold may include any suitable inequality, such as (but not limited to) greater than, greater than or equal to, less than, or less than or equal to.

[0071] Although sparse vector indices 404a to 404d have been described as equivalent to binary mode selection indicators by comparing them with a threshold to determine subset 406, this is not intended to be limiting. In embodiments, sparse vector indices 404a to 404d can act as weighting vectors, which are provided to the defect detection model 104 during training. In this respect, the defect detection model 104 can take into account the weights of sparse vector indices 404a to 404d when using images 128a to 128d during training, such that all modes in the dataset 126 can be used to train the defect detection model 104. For example, the 4D tensor channels can be weighted based on sparse vector indices 404a to 404d during training of the defect detection model 104.

[0072] Next, the sparse vector indices 404a to 404d can be updated based on defect detection tests (e.g., step 210 of performing defect detection tests). If defect detection model 104a has improved stability or sensitivity compared to defect detection models 104b or 104c, then the sparse vector indices used to train defect detection model 104a can be increased. Similarly, if defect detection model 104a has a reduced defect detection capability, the sparse vector indices can be reduced. Therefore, the sparse vector indices 404a to 404d can be updated according to a selected sparsity optimization technique (e.g., L1 or L2). α (Optimization constraints) are used to optimize the sensitivity and / or stability of the defect detection model 104.

[0073] When the pattern selection model 103 contains a sparse vector 402, the runtime pattern can optionally be selected by first applying a threshold to the dataset. In this regard, indices 404a to 404d of the sparse vector 402 that exceed the threshold can be considered the most relevant patterns. In some embodiments, the threshold can directly determine the runtime pattern (e.g., without further application of a defect detection model to generate a ranking table). This can occur when the index of the sparse vector 402 is high (e.g., close to 1) (indicating that the associated optical pattern is the best optical pattern); or when all but a few (e.g., one or two) of the sparse vector indices are low (e.g., close to 0). In other embodiments, the threshold can determine a plurality of most relevant patterns (e.g., ten or more) indicating that any one of the plurality of most relevant patterns is the best optical pattern, in which case a ranking table can be generated, allowing the user to select a runtime pattern from the plurality of most relevant patterns.

[0074] Figure 5 Step 208: Describing a training defect detection model 104 according to one or more embodiments of the present disclosure.

[0075] In this embodiment, the defect detection model 104 is trained using a neural network. The neural network may include at least one of a deep generative model, a convolutional neural network (CNN), a generative adversarial network (GAN), a conditional generative adversarial network (cGAN), a variational autoencoder (VAE), a representation learning network, or a transformer model (e.g., a bidirectional encoder representation from a transformer (BERT); a generative pre-trained transformer (GPT), etc.).

[0076] The neural network may comprise multiple subnetworks trained on subsets of images 502a to 502c associated with a selected optical mode. Such subsets of images 502a to 502c may be selected by a mode selection model based on at least one of a random channel dropout vector 302, a sparse vector 402, or a model-agnostic meta-learning algorithm. Each subset of images 502a to 502c may contain four-dimensional (4D) tensor data (or other arrays of this dimension). The tensor data may contain any suitable format (e.g., (N, W, H, C) format) corresponding to the number of images in a batch, the width of the images, the height of the images, and the number of channels in the images (e.g., channels for grayscale, three channels for red, green, and blue (RGB), etc.). Given the 4D tensor data, the neural network may train a defect detection model 104 to detect defects.

[0077] Systems and methods for incorporating neural networks and forward physics models for semiconductor applications are described in U.S. Patent No. 10,346,740 by Jing Zhang, the entire contents of which are incorporated herein by reference. Diagnostic systems and methods for configuring deep learning models for semiconductor applications are described in U.S. Patent Application No. 15 / 694,719, published by Jing Zhang as US 2018 / 0107928, the entire contents of which are incorporated herein by reference.

[0078] In this embodiment, the defect detection model 104 can be trained repeatedly. The number of repetitions may correspond to the number of times the defect detection model 104 is trained using a combination of optical patterns. In this regard, each pattern in the dataset 126 can be provided to the defect detection model for learning over a given number of repetitions.

[0079] Overall Reference Figures 1A to 5 The system 100 and method 200 are described in further detail.

[0080] In an embodiment, system 100 is configured to perform pattern selection model 103 using both random channel drop vector 302 and sparse vector 402. For example, system 100 may use sparse vector 402 to determine a first subset of relevant patterns. Then, system 100 may be configured to apply random channel drop vector 302 to further limit the number of relevant patterns.

[0081] Although the pattern selection model 103 has been described as containing either a sparse vector 302 or a random channel drop-off vector 402, this is not intended to be limiting. In this respect, the pattern selection model 103 can be configured to select a subset of the training data via model-agnostic meta-learning (MAML). MAML is described in Chelsea Finn’s “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks,” the entire contents of which are incorporated herein by reference.

[0082] The pattern selection model 103 may also include a stepwise regression technique. For example, the pattern selection model 103 may include forward selection. Through forward selection, a first optical pattern can be selected, wherein this first optical pattern has the desired quality (e.g., the defect detection model 104 has the desired stability and / or sensitivity when trained using the optical pattern). Additional optical patterns can be selected, which further increase the stability or sensitivity of the defect detection model. As another example, the pattern selection model 103 may include backward selection. Backward selection can begin training the defect detection model using all optical patterns. The first optical pattern can be removed, the removal of which causes a minimum decrease in the stability or sensitivity of the defect detection model. Forward selection or backward selection can be repeated or until the desired number of optical patterns are determined.

[0083] While the pattern selection model 103 can be configured to determine runtime optical patterns from every possible combination of patterns in the verification subsystem 102, this is generally not feasible under time constraints. Therefore, it is desirable to limit the pattern selection model 103 to configuring a number of candidate patterns for the verification subsystem 102. In embodiments, a list of candidate patterns is determined (e.g., by one or more of the user or controller 101). The list of candidate optical patterns can be determined in part by dimensionality reduction analysis. In this regard, one or more redundant patterns can be removed. For example, candidate patterns can be determined by correlation analysis or principal component analysis. Correlation analysis can be performed by calculating the cross-correlation between pattern pairs (e.g., Pearson correction coefficients). Any pattern pairs with high coefficients (redundant patterns) are removed. Principal component analysis (PCA) can be performed by retaining high-variant principal components. The number and configuration of patterns described herein are not intended to be limiting.

[0084] In an embodiment, method 200 includes region of interest optimization. Through region of interest optimization, a region of the sample can be selected as the region of interest, having one or more suitable defects for joint optical selection and defect detection. One or more additional images can be obtained from the region of interest to train the defect detection model. Design- and noise-based regions of interest are described in U.S. Patent Application No. 16 / 364,161, published by Brian Duffy as US 2020 / 0126212, the entire contents of which are incorporated herein by reference.

[0085] All methods described herein may include storing the results of one or more steps of the method embodiments in memory. The results may include any of the results described herein and may be stored in any manner known in the art. The memory may include any memory described herein or any other suitable storage medium known in the art. After the results have been stored, they may be accessed in memory and used by any of the method or system embodiments described herein, formatted for display to a user, used by another software module, method, or system, and the like. Furthermore, the results may be stored “permanently,” “semi-permanently,” “temporarily,” or for a period of time. For example, the memory may be random access memory (RAM), and the results may not necessarily be retained in memory indefinitely.

[0086] It should be further considered that each of the embodiments of the methods described above may include any other step of any other method described herein. Furthermore, each of the embodiments of the methods described above may be performed by any of the systems described herein.

[0087] As used throughout this disclosure, the term "sample" generally refers to a substrate formed of a semiconductor or non-semiconductor material (e.g., a wafer or the like). For example, semiconductor or non-semiconductor materials may include (but are not limited to) single-crystal silicon, gallium arsenide, and indium phosphide. A sample may comprise one or more layers. For example, such layers may include (but are not limited to) photoresist (including photoresist), dielectric materials, conductive materials, and semiconductor materials. Many different types of such layers are known in the art, and the term "sample" as used herein is intended to cover a sample on which all types of such layers can be formed. The one or more layers formed on the sample may be patterned or unpatterned. For example, a sample may comprise multiple dies, each having repeatably patterned features. The formation and processing of such material layers can ultimately result in a completed device. Many different types of devices can be formed on a sample, and the term "sample" as used herein is intended to cover a sample on which any type of device known in the art is being fabricated. Furthermore, for the purposes of this disclosure, the terms "sample" and "wafer" should be interpreted as interchangeable. Furthermore, for the purposes of this disclosure, the terms patterning device, mask, and photomask should be interpreted as interchangeable.

[0088] Those skilled in the art will recognize that, for clarity of concept, the component operations, devices, objects, and accompanying discussions described herein are used as examples, and various configuration modifications are taken into account. Therefore, as used herein, the specific examples illustrated and the accompanying discussions are intended to represent their more general categories. In general, the use of any particular example is intended to represent its category, and the absence of specific components, operations, devices, and objects should not be considered limiting.

[0089] Regarding the use of any generally plural and / or singular terms in this document, those skilled in the art can convert them from plural to singular and / or from singular to plural depending on the context and / or application. For clarity, various singular / plural arrangements are not explicitly described in this document.

[0090] The topics described herein sometimes illustrate different components contained within or connected to other components. It should be understood that such depicted architectures are merely illustrative, and many other architectures can in fact be implemented to achieve the same functionality. Conceptually, any arrangement of components used to achieve the same functionality is effectively “associated” to achieve the desired functionality. Therefore, any two components combined herein to achieve a particular functionality can be considered “associated” with each other to achieve the desired functionality, regardless of the architecture or intermediate components. Similarly, any two such associated components can also be considered “connected” or “coupled” to each other to achieve the desired functionality, and any two components that can be suchly associated can also be considered “coupleable” to each other to achieve the desired functionality. Specific instances of coupleability include (but are not limited to) components that can be physically mated and / or physically interact, and / or components that can wirelessly interact and / or wirelessly interact, and / or components that can logically interact and / or logically interact.

[0091] Furthermore, it should be understood that the invention is defined by the appended claims. Those skilled in the art will understand that, generally, the terms used herein, and especially in the appended claims (e.g., the body of the appended claims), are intended to be “open-ended” terms (e.g., the term “comprising” should be interpreted as “comprising (but not limited to)”, the term “having” should be interpreted as “at least having”, the term “including” should be interpreted as “including (but not limited to)”, and the like). Those skilled in the art will further understand that if a particular number of claims is desired to be introduced, this intention will be explicitly stated in the claims, and if such a statement is lacking, then this intention does not exist. For example, as an aid to understanding, the appended claims may contain the use of the introductory phrases “at least one” and “one or more” to introduce the claims. However, the use of such phrases should not be construed as implying that the introduction of a claim statement by the indefinite article “a(a)” or “an(an)” limits any particular claim containing such an introductory claim statement to the invention containing only that statement, even if the same claim contains the introductory phrase “one or more” or “at least one” and indefinite articles such as “a(a)” or “an(an)” (e.g., “a(a)” and / or “an(an)” should generally be interpreted as meaning “at least one” or “one or more”); the same applies to the use of definite articles used to introduce a claim statement. Furthermore, even if a specific number of claim statements is explicitly stated, those skilled in the art will recognize that such a statement should generally be interpreted as meaning at least the number of stated statements (e.g., a bare statement of “two statements” (without other modifiers) generally means at least two statements or two or more statements). Furthermore, in examples where a convention similar to "at least one of A, B, and C, and similar ones" is used, this construction is generally intended to be understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" would include (but is not limited to) systems having only A, only B, only C, both A and B, both A and C, both B and C, and / or both A, B, and C, and similar ones). In examples where a convention similar to "at least one of A, B, or C, and similar ones" is used, this construction is generally intended to be understood by a person skilled in the art (e.g., "a system having at least one of A, B, or C" would include (but is not limited to) systems having only A, only B, only C, both A and B, both A and C, both B and C, and / or both A, B, and C, and similar ones). Those skilled in the art will further understand that virtually any transitional conjunction and / or phrase presenting two or more alternative items, whether in the specification, claims, or drawings, should be understood to include the possibility that one, any, or both of the items are included.For example, the phrase “A or B” would be understood as including the possibility of “A” or “B” or “A and B”.

[0092] It will be understood from the foregoing description that this disclosure and its many accompanying advantages are readily apparent, and that various changes can be made to the form, construction, and arrangement of the components without departing from the disclosed subject matter or sacrificing all its material advantages. The forms described are for illustrative purposes only, and the appended claims are intended to cover and encompass such changes. Furthermore, it should be understood that the invention is defined by the appended claims.

Claims

1. A system comprising: A controller, communicatively coupled to an inspection subsystem configured to image at least one sample when configured to have any of a plurality of candidate optical modes, the controller comprising one or more processors configured to execute program instructions that cause the one or more processors to repeatedly and jointly perform optical mode selection and defect detection training: Receive defect data of at least one defect on at least a portion of the at least one sample; The inspection subsystem receives at least one image and stores the at least one image in a dataset, wherein the at least one image is associated with at least one defect detected on at least one portion of the at least one sample by the inspection subsystem configured to have candidate optical modes from the plurality of candidate optical modes; The optical mode is selected from the plurality of candidate optical modes by executing the mode selection model; and The defect detection model is trained using images associated with one or more selected optical modes. The one or more processors are further configured to determine at least one runtime optical mode from the plurality of candidate optical modes.

2. The system according to claim 1, wherein the mode selection model comprises: A sparse vector containing multiple indices, each of which contains a mode selection weight between zero and a first value.

3. The system of claim 2, wherein the one or more optical modes selected by the mode selection model are selected by applying a threshold to the sparse vector.

4. The system of claim 2, wherein the one or more optical modes selected by the mode selection model are selected by providing the plurality of indices as weighted vectors to the defect detection model.

5. The system according to claim 1, wherein the mode selection model comprises: A random channel drop vector, which contains multiple indices, each of which contains a zero or non-zero mode selection weight.

6. The system of claim 5, wherein the plurality of indices are randomly set to zero or a non-zero value during each iteration.

7. The system of claim 6, wherein the one or more optical modes of the dataset used to train the defect detection model are determined by the plurality of indices having the non-zero values.

8. The system according to claim 1, wherein the mode selection model includes a model-agnostic meta-learning algorithm.

9. The system according to claim 1, wherein the mode selection model includes at least one of a forward selection algorithm or a backward selection algorithm.

10. The system of claim 1, wherein determining the at least one runtime mode comprises generating a ranking table, the ranking table comprising at least one of signal-to-noise ratio, receiver operating characteristic curve, capture rate, interference rate, or computational cost.

11. The system of claim 1, wherein the plurality of candidate optical modes are determined by dimensionality reduction, the dimensionality reduction comprising at least one of correlation analysis or principal component analysis.

12. The system of claim 1, wherein each of the plurality of candidate optical modes comprises wavelength, focal length, aperture, and bandwidth.

13. The system according to claim 1, wherein the defect detection model comprises at least one of a deep generative model, a convolutional neural network, a generative adversarial network, a conditional generative adversarial network, a variational autoencoder, a representation learning network, or a transformer model.

14. The system of claim 1, wherein the inspection subsystem comprises a broadband plasma inspection tool.

15. The system of claim 1, further comprising performing defect inspection tests.

16. The system of claim 15, further comprising using the defect detection model to perform inference to evaluate at least one of the stability or sensitivity of the defect detection model.

17. The system of claim 1, wherein when the one or more optical modes do not contain associated images in the dataset, the verification subsystem is configured to image the at least one sample when configured to have the one or more optical modes selected from the plurality of candidate modes by the mode selection model.

18. A method for performing optical mode selection and defect detection training, comprising: Receive defect data of at least one defect on at least a portion of at least one sample; At least one image is received from the inspection subsystem and stored in the dataset, wherein the at least one image is associated with at least one defect detected on at least one portion of the at least one sample by the inspection subsystem configured to have candidate optical modes among a plurality of candidate optical modes; The optical mode is selected from the plurality of candidate optical modes by executing the mode selection model; The defect detection model is trained using the images associated with one or more optical modes selected by the mode selection model. and Perform defect inspection tests.

19. The method of claim 18, wherein the mode selection model comprises one or more of the following: random channel drop vector, sparse vector, model agnostic meta-learning algorithm, forward selection algorithm, or backward selection algorithm.

20. A system comprising: A verification subsystem configured to image at least one sample when configured to have multiple candidate optical modes; A controller, communicatively coupled to the inspection subsystem, includes one or more processors configured to execute program instructions that cause the one or more processors to repeatedly and jointly perform optical mode selection and defect detection training. Receive defect data of at least one defect on at least a portion of the at least one sample; The inspection subsystem receives at least one image and stores the at least one image in a dataset, wherein the at least one image is associated with at least one defect detected on at least one portion of the at least one sample by the inspection subsystem configured to have candidate optical modes from the plurality of candidate optical modes; The optical mode is selected from the plurality of candidate optical modes by executing the mode selection model; and The defect detection model is trained using images associated with one or more selected optical modes. The one or more processors are further configured to determine at least one runtime optical mode from the plurality of candidate optical modes.

Images