Machine learning based inspection of semiconductor samples and training thereof

Abstract

A system and method of runtime inspection of a semiconductor sample is provided. The method includes obtaining a runtime image representing an inspection region of the sample, the runtime image having a relatively low signal-to-noise ratio (SNR), and processing the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model is previously trained for the given inspection application using one or more training samples, each training sample representing a respective reference region sharing a same design pattern as the inspection region and including: a first training image of the respective reference region having a relatively low SNR, and label data indicating ground truth in the respective reference region regarding the given inspection application, the label data obtained by labeling a second training image of the respective reference region having a relatively high SNR.

Description

TECHNICAL FIELD

[0001] The presently disclosed subject matter relates generally to the field of inspection of semiconductor samples, and more specifically, to inspection of samples using a specially trained machine learning model. BACKGROUND

[0002] Current demands for high density and high performance associated with very large scale integration of fabricated devices require sub-micron features, increased transistor and circuit speed, and improved reliability. As semiconductor processes develop, pattern dimensions, such as line widths, and other types of critical dimensions are continually shrinking. Such demands require the formation of device features with high precision and high uniformity, which in turn necessitates careful monitoring of the manufacturing process, including automated inspection of the devices while the devices are still in the form of semiconductor wafers.

[0003] Runtime inspection can generally employ a two-stage process, e.g., inspecting a sample, followed by reviewing sampled locations of potential defects. Inspection typically involves generating some output (e.g., image, signal, etc.) for a sample by directing light or electrons to the wafer and detecting light or electrons from the wafer. During the first stage, the surface of the sample is inspected at high speed and at relatively low resolution. Defect detection is typically performed by applying a defect detection algorithm to the inspection output. A defect map is produced to show suspicious locations on the sample where there is a high probability of defects. During the second stage, at least some of the suspicious locations are more thoroughly analyzed at relatively high resolution to determine different parameters of the defects, such as class, thickness, roughness, size, etc.

[0004] Inspection can be provided by using non-destructive inspection tools during or after the manufacturing of the sample to be inspected. As non-limiting examples, various non-destructive inspection tools include scanning electron microscopes, atomic force microscopes, optical inspection tools, etc. In some cases, the two stages can be implemented by the same inspection tool, and in some other cases, the two stages are implemented by different inspection tools.

[0005] An inspection process can include multiple inspection steps. A manufacturing process of semiconductor devices can include various processes, such as etching, deposition, planarization, growth (such as epitaxial growth), implantation, etc. Inspection steps can be performed several times, e.g., after certain process procedures and / or after manufacturing of certain layers, etc. Additionally or alternatively, each inspection step can be repeated multiple times, e.g., for different wafer locations or for the same wafer location under different inspection settings.

[0006] For example, inspection processes are used at various steps during semiconductor manufacturing to detect and classify defects on a sample, as well as to perform metrology-related operations. The effectiveness of the inspection can be improved through automation of the process(es), such as, for example, defect detection, automated defect classification (ADC), automated defect review (ADR), image segmentation, automated metrology-related operations, etc. Automated inspection systems ensure that the manufactured parts meet the expected quality standards and provide useful information about adjustments that can need to be made to the manufacturing tools, equipment, and / or composition depending on the type of defects identified.

[0007] In some cases, machine learning techniques can be used to assist the automated inspection process in order to facilitate higher throughput. For example, supervised machine learning can be used to implement an accurate and efficient solution for automating a particular inspection application based on fully-labeled training images. SUMMARY

[0008] According to certain aspects of the presently disclosed subject matter, there is provided a computerized system of runtime inspection of a semiconductor sample, the system comprising processing and memory circuitry (PMC) configured to: obtain a runtime image representing an inspection region of a semiconductor sample, the runtime image having a relatively low signal-to-noise ratio (SNR); and process the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model is previously trained for the given inspection application using one or more training samples, each training sample representing a respective reference region sharing a same design pattern as the inspection region and comprising: a first training image of the respective reference region, the first training image having a relatively low SNR similar to the low SNR of the runtime image; and label data indicating ground truth about the given inspection application in the respective reference region, the label data being obtained by labeling a second training image of the respective reference region having a relatively high SNR.

[0009] In addition to the features above, the system according to this aspect of the presently disclosed subject matter can include one or more of the following listed features (i) to (xii) in any desired combination or arrangement technically possible:

[0010] (i). The runtime image and the one or more training samples are acquired by an e-beam tool.

[0011] (ii). The given inspection application is one of: a segmentation application for segmenting the runtime image into one or more segments in the inspection region, a metrology application for obtaining one or more measurement values about structural elements in the inspection region, a defect detection application for detecting one or more candidate defects in the inspection region, and a defect classification application for classifying one or more defects in the inspection region.

[0012] (iii). generating a first training image based on a first number of training frames acquired for the reference region, and generating a second training image based on a second number of training frames acquired for the reference region. The first number is less than the second number, and the runtime image is generated based on the first number of runtime frames.

[0013] (iv). the second number of training frames includes the first number of training frames.

[0014] (v). generating the first training image based on a first dose of electrons, and generating the second training image based on a second dose of electrons. The first dose is less than the second dose, and the runtime image is generated based on the first dose of electrons.

[0015] (vi). the label data is obtained based on at least one of: manual annotation, synthetically generated labels based on design data, machine learning derived labels, or a combination thereof.

[0016] (vii). the first training image and the second training image are registered to correct for an offset between them so that the first training image and the label data are aligned.

[0017] (viii). the first training image is acquired before the second training image.

[0018] (ix). the inspection region is a portion of a sensitive layer of the semiconductor sample that is inspectable only via an image having a relatively low SNR.

[0019] (x). the given inspection application is a segmentation application for segmenting the runtime image into one or more segments in the inspection region, and the label data indicates a particular segment to which each pixel of at least a portion of the runtime image in the one or more segments belongs.

[0020] (xi). the inspection region is from an inspection die of the semiconductor sample, and the respective reference region is from a reference die of the semiconductor sample or a reference die of a different semiconductor sample.

[0021] (xii). the relatively low SNR and the relatively high SNR are relative to each other, or are relative to a threshold.

[0022] According to other aspects of the presently disclosed subject matter, there is provided a method of runtime inspection of a semiconductor sample, the method performed by a processor and memory circuitry (PMC) and comprising: obtaining a runtime image representative of an inspection region of the semiconductor sample, the runtime image having a relatively low signal-to-noise ratio (SNR); and processing the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model is previously trained for the given inspection application using one or more training samples, each training sample being representative of a respective reference region sharing a same design pattern as the inspection region and comprising: a first training image of the respective reference region, the first training image having a relatively low SNR similar to the low SNR of the runtime image; and label data indicative of ground truth in the respective reference region pertaining to the given inspection application, the label data obtained by labeling a second training image of the respective reference region having a relatively high SNR.

[0023] According to other aspects of the presently disclosed subject matter, there is provided a method of training a machine learning model usable for inspecting a semiconductor sample, the method performed by a processor and memory circuitry (PMC) and comprising: obtaining a training set comprising one or more training samples, each training sample being representative of a respective reference region having a given design pattern, the training sample comprising: a first training image having a relatively low signal-to-noise ratio (SNR); and label data indicative of ground truth in the respective reference region pertaining to a given inspection application, the label data obtained by labeling a second training image of the respective reference region having a relatively high SNR; and training a ML model using the training set for the given inspection application; wherein the ML model, upon being trained, is usable to process a runtime image representative of an inspection region of a sample sharing a same design pattern as the given design pattern and to obtain inspection data specific to the given inspection application, the runtime image having a relatively low SNR similar to the low SNR of the first training image.

[0024] These aspects of the disclosed subject matter can include one or more of features (i) to (xii) listed above with respect to the system, mutatis mutandis, in any desired combination or arrangement technically possible.

[0025] According to other aspects of the presently disclosed subject matter, there is provided a non-transitory computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform a method for runtime inspection of a semiconductor sample, the method comprising: obtaining a runtime image representing an inspection region of the semiconductor sample, the runtime image having a relatively low signal-to-noise ratio (SNR); and processing the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model is previously trained for the given inspection application using one or more training samples, each training sample representing a respective reference region sharing a same design pattern as the inspection region and comprising: a first training image of the respective reference region, the first training image having a relatively low SNR similar to the low SNR of the runtime image; and label data indicating ground truth in the respective reference region pertaining to the given inspection application, the label data being obtained by labeling a second training image of the respective reference region having a relatively high SNR.

[0026] According to other aspects of the presently disclosed subject matter, there is provided a non-transitory computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform a method for training a machine learning model usable for inspecting a semiconductor sample, the method comprising: obtaining a training set comprising one or more training samples, each training sample representing a respective reference region having a given design pattern, the training sample comprising: a first training image having a relatively low signal-to-noise ratio (SNR); and label data indicating ground truth in the respective reference region pertaining to a given inspection application, the label data being obtained by labeling a second training image of the respective reference region having a relatively high SNR; and training a ML model using the training set for the given inspection application; wherein the ML model, upon being trained, is usable to process a runtime image representing an inspection region of a sample sharing a same design pattern as the given design pattern and to obtain inspection data specific to the given inspection application, the runtime image having a relatively low SNR similar to the low SNR of the first training image.

[0027] These aspects of the disclosed subject matter can include one or more of features (i) through (xii) listed above with respect to the system, mutatis mutandis, in any desired combination or arrangement technically possible. BRIEF DESCRIPTION OF DRAWINGS

[0028] For the purposes of the present disclosure and to understand how the same can be put into effect, embodiments will be described, by way of non-limiting examples only, with reference to the accompanying drawings in which:

[0029] Figure 1A generalized block diagram of an inspection system in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0030] Figure 2 A generalized flowchart of training a machine learning model that can be used to inspect a semiconductor sample in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0031] Figure 3 A generalized flowchart of runtime inspection of a semiconductor sample using a trained ML in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0032] Figure 4 A generalized flowchart of generating a training set for training an ML model that can be used to inspect a semiconductor sample in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0033] Figure 5 A schematic diagram of a training process of an ML model in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0034] Figure 6 An example of a first training image in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0035] Figure 7 A schematic diagram of a runtime inspection process using an ML model in accordance with certain embodiments of the presently disclosed subject matter. DETAILED DESCRIPTION

[0036] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent to one skilled in the art that the presently disclosed subject matter can be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the presently disclosed subject matter.

[0037] Unless specifically stated otherwise, as apparent from the following discussions, it is appreciated that throughout the specification discussions utilizing terms such as "obtaining," "processing," "training," "acquiring," "segmenting," "detecting," "classifying," "generating," "registering," or the like, refer to the action(s) and / or process(es) of a computer that manipulate and / or transform data into other data, the data represented as physical (such as electronic) quantities within the computer's registers and / or memories, and / or the data representing a physical object. The term "computer" should be expansively construed to include any kind of hardware-based electronic device with data processing capabilities, including, by way of non-limiting examples, the inspection system, the training system, and their respective parts disclosed in the present application, as the action(s) of a computer are merely provided as an example for the action(s) and / or process(es) of a hardware-based electronic device.

[0038] The terms “non-transitory storage” and “non-transitory storage media” as used herein should be broadly interpreted to encompass any volatile or non-volatile computer storage media, suitable for the current disclosure. The terms should be construed to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) that store one or more sets of instructions that, when executed by a computer, cause the computer to perform any one or more of the methods of the present disclosure. The terms should therefore include, but are not limited to, random access memory (“RAM”), read-only memory (“ROM”), magnetic storage, optical storage, flash memory devices, and the like.

[0039] The term “sample” as used in the present specification should be broadly interpreted to encompass any kind of physical object or substrate used in the fabrication of semiconductor integrated circuits, magnetic heads, flat panel displays, and other semiconductor articles, including wafers, masks, reticles, and other structures, combinations of the above, and / or portions thereof. The sample is also referred to herein as a semiconductor sample, and can be produced by manufacturing equipment performing a corresponding manufacturing process.

[0040] The term “inspection” as used in the present specification should be broadly interpreted to encompass any type of operational process related to defect detection, defect review, and / or various types of defect classification, segmentation, and / or metrology operations during and / or after the fabrication of a sample. Inspection is provided by using non-destructive inspection tools during or after the fabrication of a sample to be inspected. As non-limiting examples, an inspection process can include runtime scanning (in a single scan or in multiple scans), imaging, sampling, detection, review, measurement, classification, and / or other operations provided for a sample or portions thereof using the same or different inspection tools. Likewise, inspection can be provided prior to the fabrication of a sample to be inspected, and can include, for example, generating inspection recipe(s) and / or other setup operations. Note that the term “inspection” or its derivatives as used in the present specification are not limited in resolution or size of the inspection area unless specifically stated otherwise. As non-limiting examples, a variety of non-destructive inspection tools include scanning electron microscopes (SEMs), atomic force microscopes (AFMs), optical inspection tools, and the like.

[0041] The term "metrology operation" as used in the present specification should be broadly interpreted to encompass any metrology operation procedure for extracting metrology information related to one or more structural elements on a semiconductor sample. In some embodiments, the metrology operation can include a measurement operation, such as, for example, a critical dimension (CD) measurement performed with respect to certain structural elements on the sample, including, but not limited to, the following: dimensions (e.g., line width, line spacing, contact diameter, element size, edge roughness, gray scale statistics, etc.), element shape, intra-element or inter-element distances, relevant angles, overlay information related to elements corresponding to different design levels, etc. For example, by employing image processing techniques to analyze the measurement results, such as measured images. It is noted that the term "metrology" or its derivatives as used in the present specification is not limited by the measurement technique, measurement resolution, or size of the inspection area, unless specifically stated otherwise.

[0042] The term "defect" as used in the present specification should be broadly interpreted to encompass any kind of abnormality or undesirable feature / function formed on a sample. In some cases, the defect can be a defect of interest (DOI), which is a real defect that has some impact on the functionality of the manufactured device, and thus it is in the customer's interest to detect such defects. For example, any "killer" defect that can cause yield loss can be indicated as a DOI. In other cases, the defect can be a nuisance (also referred to as a "false positive" defect), which can be negligible since it has no impact on the functionality of the completed device and also does not affect yield.

[0043] The term "candidate defect" as used in the present specification should be broadly interpreted to encompass a suspicious defect location on a sample that is detected as having a relatively high probability of being a defect of interest (DOI). Thus, upon review, the candidate defect can actually be a DOI, or in some other cases, it can be a nuisance as described above, or can be random noise resulting from different variations (e.g., process variations, color variations, mechanical and electrical variations, etc.) during inspection.

[0044] The term "design data" as used in the present specification should be broadly interpreted to encompass any data indicative of the layered physical design (layout) of a sample. The design data can be provided by the respective designer and / or can be derived from the physical design (e.g., by complex simulations, simple geometrical and Boolean operations, etc.). The design data can be provided in different formats, such as, by way of non-limiting examples, GDSII format, OASIS format, etc. The design data can be presented in a vector format, a gray scale intensity image format, or otherwise.

[0045] It will be appreciated that certain features of the presently disclosed subject matter described in the context of separate embodiments can also be provided in combination, unless otherwise specifically stated. Conversely, various features of the presently disclosed subject matter described in the context of a single embodiment can also be provided separately or in any appropriate subcombination. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the methods and devices. For the purpose of clarity, the description will not list every possible feature of the presently disclosed subject matter.

[0046] In view of the foregoing, attention is directed to Figure 1 , Figure 1 A functional block diagram of an inspection system in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0047] As part of a sample manufacturing process, images can be obtained using Figure 1 A semiconductor sample (e.g., a wafer, a die, or a portion thereof) is inspected using the inspection system 100 shown. As noted above, inspection referred to herein can be interpreted to encompass any kind of operation related to defect review / detection, various types of defect classification, segmentation, and / or metrology operations, such as, for example, critical dimension (CD) measurements with respect to a sample. The inspection system 100 shown includes a computer-based system 101 capable of enabling automated inspection of a semiconductor sample based on machine learning (ML). In accordance with certain embodiments of the presently disclosed subject matter, the system 101 can be configured to use a trained machine learning (ML) model to inspect a semiconductor sample on-the-fly based on images obtained during sample manufacturing (also referred to herein as manufacturing process (FP) images or runtime images). In some embodiments, the system 101 can be configured as a training system capable of training an ML model using a specially generated training set during a training / setup phase.

[0048] The system 101 is operatively connected to one or more inspection tools 120. The inspection tools 120 are configured to capture runtime images and / or training images, process the captured images, and / or provide measurements related to the captured images.

[0049] For example, as used herein, a runtime image and / or a training image can refer to an original image of a sample captured during a manufacturing process, a derivative of the captured image obtained through various pre-processing stages, and / or an image based on computer generated design data. For example, the images can be selected from, for example, images of portions of a sample captured by a scanning electron microscope (SEM) or an optical inspection system, SEM images centered approximately on a defect to be classified by the ADC, SEM images of a larger area in which a defect is to be located by the ADR, registered images of different inspection modalities corresponding to the same location, segmented images or height map images, etc. Note that in some cases, the images referred to herein can include image data (e.g., captured images, processed images, etc.) and associated digital data (e.g., metadata, handcrafted attributes, etc.). Note further that the image data can include data related to a layer of interest and / or to one or more layers of a sample.

[0050] As used herein, the term “inspection tool(s)” should be interpreted broadly to encompass any tool that can be used to inspect a relevant process, including, by way of non-limiting examples, imaging, scanning (in single or multiple scans), sampling, review, measurement, classification, and / or other processes provided with respect to a sample or portions thereof. The inspection tool(s) 120 can include one or more review tools and / or one or more inspection tools. In some cases, at least one of the inspection tools 120 can be an inspection tool configured to scan a sample (e.g., an entire wafer, an entire die, or portions thereof) to capture inspection images (typically at a relatively high speed and / or low resolution) to detect potential defects (i.e., candidate defects). In some cases, at least one of the inspection tools 120 can be a review tool configured to capture review images of at least some of the candidate defects detected by the inspection tool to determine whether the candidate defects are indeed defects of interest (DOIs). Such review tools are typically configured to inspect one specimen at a time (typically at a relatively low speed and / or high resolution). The inspection tool and the review tool can be different tools located at the same or different locations, or a single tool operating in two different modes. In some cases, at least one inspection tool can have metrology capabilities and can be configured to perform metrology operations on the images.

[0051] Without limiting the scope of the disclosure in any way, it should also be noted that the inspection tool 120 can be implemented as various types of inspection machines, such as optical inspection machines, electron beam inspection machines (e.g., scanning electron microscopes (SEMs)), atomic force microscopes (AFMs), or transmission electron microscopes (TEMs), etc. In some cases, the same inspection tool can provide both low resolution image data and high resolution image data. The resulting image data (low resolution image data and / or high resolution image data) can be transmitted to the system 101 directly or via one or more intermediate systems. The present disclosure is not limited to any particular type of inspection tool and / or resolution of image data produced by the inspection tool.

[0052] As previously mentioned, the inspection system 100 includes a computer-based system 101 that is capable of runtime inspection of semiconductor samples based on runtime images obtained during sample manufacturing using a trained machine learning (ML) model. In order to train the ML model, sufficient training data is required. However, during the setup phase, training images of the samples to be inspected can not be available. In some cases, the training images of the test / reference wafers used can not include sufficient variation in the structure parameters to produce a robust, well-trained accurate model with respect to process variations in actual production. Therefore, in order to train the ML to achieve accurate inspection, training of the ML model is typically performed at the customer site using images acquired from actual production wafers. For example, SEM images of production wafers can be captured in the FAB and used for the purpose of training the ML model.

[0053] A SEM image can be generated by aggregating a series of frames captured for a region of a semiconductor sample, the frames being sequentially acquired by an electron beam inspection tool such as a scanning electron microscope (SEM). In order to obtain a higher quality SEM image (e.g., higher signal-to-noise ratio (SNR)), the electron beam dose used to scan the wafer (which can be reflected as the number of frames captured for the region and / or the intensity of the electron beam used to capture the frames) must reach a certain amount in order to reduce the noise in the resulting image.

[0054] On the other hand, due to the impact of the electron beam hitting on the sample, the sample can be physically damaged. This phenomenon is referred to as "shrinkage" or "thinning". For example, depending on different layers and / or materials of the sample, a typical shrinkage magnitude can be between 5% to 10% of the size of the structural features on the wafer, which is undesirable for the customer. In some cases, to reduce the damage to the production wafer, particularly certain sensitive layers of the wafer, the electron dose of the electron beam tool used to acquire the frames is limited. However, the SEM images obtained under such limitation are typically very noisy with low SNR. Performing annotation on low SNR images is undesirable and in some cases even impossible, and is very challenging even for a user to perform manual annotation. The resulting annotation on such images can be inaccurate and error-prone. Since the labeled data of the annotation is used as ground truth in the training of the ML model, the ML model trained using such training data is unable to provide accurate and effective inspection of the sample.

[0055] Accordingly, certain embodiments of the present disclosure subject matter propose a system (e.g., system 101) that is capable of training a ML model using training samples with a particular composition in order to address the above problems. Certain embodiments of the present disclosure use the ML model trained in this way to perform runtime inspection, as detailed below.

[0056] The system 101 includes a processor and memory circuitry (PMC) 102, which is operably connected to a hardware-based I / O interface 126. The PMC 102 is configured to provide processing required by an operating system, as referenced to Figures 2 to 4 Further detailed below, and includes a processor (not shown separately) and a memory (not shown separately). The processor of the PMC 102 can be configured to execute several functional modules in accordance with computer-readable instructions embodied on a non-transitory computer-readable memory included in the PMC. Such functional modules are referred to below as being included in the PMC.

[0057] As previously mentioned, in certain embodiments, the system 101 can be configured as a training system capable of training a ML model using training samples during a training / setup phase. In this case, the functional modules included in the PMC 102 can include a training set generator 104, a training module 106, and a machine learning model 108. The training set generator 104 can be configured to obtain a training set including one or more training samples. Each training sample represents a respective reference region having a given design pattern. The training sample includes a first training image having a relatively low signal-to-noise ratio (SNR) and label data indicating ground truth about a given inspection application in the respective reference region. The label data is obtained by annotating a second training image of the respective reference region having a relatively high SNR.

[0058] The training module 106 can be configured to train the machine learning model 108 using the training set for a given inspection application. The ML model, after training, can be used to process a runtime image representing an inspection region sharing the same design pattern as the given design pattern and obtain inspection data specific to the given inspection application. The runtime image is acquired at a relatively low SNR. Details of the training process are described below with reference to Figure 2 and Figure 4 .

[0059] According to certain embodiments, the system 101 can be configured to use the trained ML model to inspect a semiconductor sample at runtime based on a runtime image obtained during fabrication of the sample. In this case, one or more functional modules included in the PMC 102 can include the ML model 108 that has been trained as described above. The PMC 102 can be configured to obtain, via the I / O interface 126, a runtime image representing an inspection region of the semiconductor sample. The runtime image is acquired at a relatively low signal-to-noise ratio (SNR).

[0060] The trained ML model 108 is used to process the runtime image to obtain inspection data specific to the given inspection application. As described above, the ML model is trained for the given inspection application using a training set including one or more training samples. Each training sample represents a respective reference region sharing the same design pattern as the inspection region. Similarly, as described above, each training sample includes a first training image having a relatively low SNR and label data indicating ground truth about the given inspection application in the respective reference region. The label data is obtained by annotating a second training image of the respective reference region having a relatively high SNR. Details of the runtime inspection process are described below with reference to Figure 3 .

[0061] According to certain embodiments, the ML model can be trained for different inspection applications based on specific training images and label data about the respective applications. Various applications to which the present disclosure can be applicable include, but are not limited to, the following: a segmentation application for segmenting a runtime image into one or more segments in an inspection region, a metrology application for obtaining one or more measurement values about a structural element in an inspection region, a defect detection application for detecting one or more candidate defects in an inspection region, and a defect classification application for classifying one or more defects in an inspection region, etc.

[0062] The operation of the system 101, the PMC 102, and the functional modules therein are further detailed with reference to Figures 2 to 4 .

[0063] According to certain embodiments, the ML models 108 referred to herein can be implemented as various types of machine learning models such as, for example, decision trees, support vector machines (SVMs), artificial neural networks (ANN), regression models, Bayesian networks, or an ensemble / combination thereof, etc. The learning algorithm used by the ML models can be any one of the following: supervised learning, unsupervised learning, or semi-supervised learning, etc. The presently disclosed subject matter is not limited to a particular type of ML model or a particular type of learning algorithm used by the ML model.

[0064] In some embodiments, the ML models can be implemented as deep neural networks (DNNs). The DNNs can include supervised or unsupervised DNN models that include layers organized according to a respective DNN architecture. As non-limiting examples, the layers of the DNNs can be organized according to a convolutional neural network (CNN) architecture, a recurrent neural network architecture, a recursive neural network architecture, a generative adversarial network (GAN) architecture, or other architectures. Optionally, at least some of the layers can be organized into multiple DNN subnets. Each layer of the DNNs can include a plurality of basic computational elements (CEs), which are typically referred to in the art as dimensions, neurons, or nodes.

[0065] Generally, the computational elements of a given layer can be connected with the CEs of a preceding layer and / or a succeeding layer. Each connection between a CE of the preceding layer and a CE of the succeeding layer is associated with a weighting value. A given CE can receive inputs from the CEs of the preceding layer via respective connections, each given connection being associated with a weighting value that can be applied to the input of the given connection. The weighting values can determine the relative strength of the connection, and thereby the relative influence of the respective input on the output of the given CE. The given CE can be configured to compute an activation value (e.g., a weighted sum of the inputs), and further derive an output by applying an activation function to the computed activation. The activation function can be, for example, an identity function, a deterministic function (e.g., linear, sigmoid, threshold, etc.), a stochastic function, or other suitable function. The output from the given CE can be transmitted to the CEs of the succeeding layer via respective connections. Likewise, as noted above, each connection at the output of a CE can be associated with a weighting value that can be applied to the output of the CE before the output of the CE is received as an input to a CE of the succeeding layer. In addition to the weighting values, there can also be threshold values (including limiting functions) associated with the connections and the CEs.

[0066] The weighting values and / or threshold values of the deep neural network can be initially selected prior to training and / or can be further iteratively adjusted or modified during training to achieve an optimal set of weighting values and / or threshold values in the trained DNN. After each iteration, a difference between the actual output produced by the DNN module and the target output associated with the corresponding data training set can be determined. The difference can be referred to as an error value. Training can be determined to have completed when a loss / cost function indicative of the error value is less than a predetermined value or when a limited change in performance is achieved between iterations. The input data set used to adjust the weights / thresholds of the deep neural network is referred to as a training set.

[0067] It is noted that the teachings of the presently disclosed subject matter are not constrained by the specific architecture of the ML or DNN as described above.

[0068] In some cases, in addition to the system 101, the inspection system 100 can include one or more inspection modules, such as, for example, a defect detection module and / or an automated defect review module (ADR) and / or an automated defect classification module (ADC) and / or metrology related modules and / or other inspection modules that can be used to inspect semiconductor samples. The one or more inspection modules can be implemented as standalone computers, or their functionality (or at least a portion thereof) can be integrated with the inspection tool 120. In some cases, the ML model 108 can be included in the one or more inspection modules. Alternatively, the ML model 108 can be shared between the inspection modules, or, alternatively, each of the one or more inspection modules can include its own ML model 108.

[0069] According to certain embodiments, the system 101 can include a storage unit 122. The storage unit 122 can be configured to store any data required for the operation of the system 101, such as, for example, data related to the input and output of the system 101, as well as intermediate processing results generated by the system 101. For example, the storage unit 122 can be configured to store the runtime images / training images produced by the inspection tool 120 and / or derivatives thereof. Accordingly, the images can be retrieved from the storage unit 122 and provided to the PMC 102 for further processing.

[0070] In some embodiments, system 101 can optionally include a computer-based graphical user interface (GUI) 124 configured to implement user-specified inputs related to system 101. For example, a user can be presented with a visual representation of a sample (e.g., via a display forming part of GUI 124), including image data of the sample. The user can be provided with options via the GUI to define certain operational parameters, such as, for example, a threshold with respect to SNR, a number of image frames to be captured, a particular inspection application, etc. For example, in some cases, a user can provide label data associated with a second training image by manually annotating on the image via GUI 124. The user can also view operational results on the GUI, such as, for example, inspection data specific to a given inspection application. In some cases, system 101 can be further configured to send inspection data to inspection tool 120 via I / O interface 126 for further processing. In some cases, system 101 can be further configured to send some of the inspection data to storage unit 122 and / or to an external system (e.g., a yield management system (YMS) of a FAB).

[0071] Those skilled in the art will readily appreciate that the teachings of the presently disclosed subject matter are not limited to Figure 1 the illustrated system; equivalent and / or modified functionality can be incorporated in another manner and can be implemented in any suitable combination of software and firmware and / or hardware.

[0072] It is noted that Figure 1 The illustrated inspection system can be implemented in a distributed computing environment, wherein Figure 1 The foregoing functional modules illustrated can be distributed across several local and / or remote devices and can be linked through a communication network. It is further noted that, in other embodiments, at least some of inspection tool 120, storage unit 122, and / or GUI 124 can be external to inspection system 100 and in data communication with system 101 via I / O interface 126. System 101 can be implemented as a standalone computer(s) used in conjunction with inspection tool and / or with additional inspection modules as described above. Alternatively, respective functionality of system 101 can be at least partially integrated with one or more inspection tools 120, thereby facilitating and enhancing the functionality of inspection tools 120 in inspection-related processes.

[0073] While not necessarily so, the operational processes of system 101 and 100 can correspond to some or all of the stages of the methods described with respect to Figures 2 to 4 Likewise, the methods described with respect to Figures 2 to 4 and their possible implementations can be implemented by system 101 and 100. It is thus noted that the methods described with respect to Figures 2 to 4The described methods discuss embodiments that can also be implemented as various embodiments of the systems 101 and 100, mutatis mutandis, and vice versa.

[0074] For illustrative purposes only, certain embodiments described below can be provided for training ML models that can be used for segmentation applications. Those skilled in the art will readily appreciate that the teachings of the presently disclosed subject matter are also applicable to various other inspection applications, such as, for example, defect detection, ADR, ADC, metrology related modules, etc.

[0075] With reference to Figure 2 , a generalized flowchart of training a machine learning model that can be used for inspecting a semiconductor sample, in accordance with certain embodiments of the presently disclosed subject matter, is shown.

[0076] The training data used for training the ML model in supervised learning typically includes one or more training samples, each training sample including a respective training image and its associated corresponding ground truth data. The ground truth data can include label data of the training image indicative of application specific information. For example, for an inspection application of image segmentation, each training sample can include a training image of a semiconductor sample and label data indicative of one or more segments in the training image.

[0077] The training image can be a “real world” image of a semiconductor sample obtained in its manufacturing process. The training image can be obtained in various ways, depending on the different inspection applications. As a non-limiting example, the image can be an inspection image obtained by inspecting the sample using one or more inspection tools for an application of detection of candidate defects. In another example, the image can be an review image obtained by inspecting the sample at a location of a candidate defect using one or more review tools for a defect review application of deciding whether a candidate defect detected by the inspection tools is indeed a defect and / or for a defect classification application of deciding on a class / type of the defect. Such review tools can be, for example, a scanning electron microscope (SEM) or the like.

[0078] The ground truth data can be obtained in various ways, such as, for example, by manual annotation, synthetic generation based on design data, machine learning based approaches, or a combination thereof, as will be detailed below with reference to Figure 4

[0079] ​As mentioned above, ML models for inspecting semiconductor samples are typically trained at customer sites using production wafer data. To reduce damage to production wafers caused by image acquisition (e.g., by SEM), the electron dose used by the e-beam tool to acquire frames should be limited (which can be represented by the number of frames captured and / or the intensity of the e-beam used). However, SEM images obtained with less electron dose are typically very noisy, with low SNR. Accurate labeling of such images is very difficult, and in some cases even impossible. Accordingly, the present disclosure proposes to address the above problem by acquiring training samples in a particular manner, as will be described below with reference to Figures 2 to 4 in detail.

[0080] As Figure 2 described, a training set can be obtained (202) (e.g., by a training set generator 104 in a PMC 102) including one or more training samples. Each training sample represents a respective reference region of a sample having a given design pattern. In some embodiments, the given design pattern can be a pattern of interest to be inspected on the sample.

[0081] In particular, a training sample includes a first training image (204) having a relatively low signal-to-noise ratio (SNR) (e.g., relatively lower with respect to a higher SNR of a second training image as described below or with respect to a threshold) and label data (206) indicating ground truth about a given inspection application in the respective reference region. The label data is obtained by labeling a second training image of the respective reference region having a relatively high SNR (e.g., relatively higher with respect to a threshold). The training set can be used to train a ML model (e.g., ML model 108) (208) (e.g., by a training module 106 in a PMC 102) for the given inspection application.

[0082] In some embodiments, when inspecting a region on a sample, a plurality of frames of the region can be sequentially acquired by an inspection tool, such as an e-beam tool, e.g., a SEM. The frames acquired by the e-beam tool are then aggregated to generate a final image, such as, e.g., a SEM image (e.g., by combining / averaging the plurality of frames to reduce noise in the resulting image). As previously mentioned, for the purpose of reducing damage to production wafers caused by image acquisition, the electron dose used by the e-beam tool to acquire frames of a production wafer should be limited. For example, the dose of electrons used can be reflected as the number of frames captured for a given region and / or the intensity of the e-beam used to capture the frames. For example, for the purpose of reducing damage, a relatively small number of frames are captured to generate a SEM image during runtime inspection of a region of a sample. As another example, an e-beam with a smaller energy level can be used to generate a SEM image. However, SEM images generated this way tend to have lower SNR and are not suitable for image labeling.

[0083] Accordingly, it is proposed that for one or more reference regions on the sample sharing the same given design pattern, two images can be captured for each reference region, including a first training image with a relatively low SNR and a second training image with a relatively high SNR. In particular, the first training image has the same / similar SNR as the SNR of the runtime image that will be captured at production time (after training and deploying the ML model in production) and will be inspected using the trained ML model. As previously mentioned, the SNR level of an image is generally related to the electronic dose used to inspect the sample and generate the image.

[0084] In particular, in some embodiments, the first training image can be generated based on a first dose of electrons and the second training image can be generated based on a second dose of electrons, where the first dose of electrons is less than the second dose. In particular, the first dose of electrons used to capture the first training image is the same as the dose of electrons used to capture the runtime image that will be inspected using the trained ML model. Note that the term “same” used herein can refer to both identical to each other and similar or highly related to each other. Various similarity metrics and algorithms can be used to determine the level of equivalence / similarity between them. Accordingly, wherever the term “same” is used, it should not be limited to exactly the same, but rather similar / equivalent to some extent.

[0085] For example, since the dose of electrons can be reflected as the number of frames used to capture the image and / or the intensity of the electron beam, in some cases, the first training image can be generated based on a first number of training frames acquired for the reference region and the second training image can be generated based on a second number of training frames acquired for the reference region, where the first number is less than the second number. Note that the number of runtime frames used to generate the runtime image is the same as the first number (the number of frames used to generate the first training image) to ensure the level of similarity in SNR between the first training image and the runtime image. In some cases, the first number (the number of frames used to generate the first training image) and the second number (the number of frames used to generate the second training image) can be defined according to the given inspection application. For example, for a segmentation application, the first number (which corresponds to the number of runtime frames used to generate the runtime image) can be determined based on the performance requirements of the application, such as, for example, the acceptable level of damage to the sample, accuracy, production volume, etc. The second number can be determined so as to produce the second training image with sufficient quality to ensure the level of annotation accuracy, while maintaining the relevance / correspondence with the first training image in terms of, for example, size, pattern, etc. For example, in some cases, the first number of training frames can be, for example, around 10 to 20 frames and the second number of training frames can be, for example, around 50 to 60 frames.

[0086] One or more reference regions sharing the same given design pattern on a sample can be identified in a variety of ways. For example, the design data of a die (or portion(s) thereof) can include various design patterns having particular geometrical structures and arrangements. In some embodiments, the design data can be received, and a plurality of design groups can be received, each design group corresponding to one or more die regions having the same design pattern. Accordingly, regions in a die corresponding to the same design pattern can be identified. In some embodiments, the inspection regions are from an inspection die of a semiconductor sample, and the reference regions can be from a reference die of the inspection die, the reference die being from the same semiconductor sample or from a different semiconductor sample.

[0087] It will be noted that similarly as defined above, a design pattern can be considered to be“the same” when the design patterns are identical, or when the design patterns are highly correlated or similar to each other. Various similarity metrics and algorithms can be applied to match and cluster similar design patterns, and the present disclosure should not be construed as being limited to any specific metric for deriving design groups. Clustering of design groups (i.e., partitioning from CAD data into multiple design groups) can be performed in advance, or as a preparatory step to the presently disclosed process by the PMC 102.

[0088] It will be noted that in some cases, the relatively low SNR and the relatively high SNR can be defined relative to a threshold, which can be a predetermined SNR level relevant to a particular inspection application (e.g., an SNR sufficient for the application to meet performance requirements regarding, for example, sensitivity, accuracy, throughput, etc.). In some cases, the relatively low SNR and the relatively high SNR can be defined relative to each other, e.g., as long as the low SNR is relatively lower than the high SNR or the high SNR is relatively higher than the low SNR. As described above with respect to determination of the first number of frames, the low SNR (which corresponds to the SNR level of the runtime images) can be defined, e.g., based on performance requirements of a particular inspection application, such as, for example, an acceptable level of damage to the sample, accuracy, throughput, etc. The high SNR can be defined to ensure a level of annotation accuracy of the second training images, while maintaining the correlation / correspondence of the second training images to the first training images in, for example, dimension, pattern, etc.

[0089] Figure 4 A generalized flowchart of generating a training set for training a ML model that can be used to inspect a semiconductor sample, in accordance with certain embodiments of the presently disclosed subject matter, is shown.

[0090] For a given reference region (of one or more reference regions), a first training image having a low SNR can be acquired (402) (e.g., by the inspection tool 120). A second training image having a high SNR can be acquired (404) (e.g., by the inspection tool 120). For example, as described above, the first training image can be generated based on a first number (N1) of training frames (e.g., 10-20 frames) acquired for the reference region, and the second training image can be generated based on a second number (N2) of training frames (which is greater than the first number, e.g., 50-60 frames) acquired for the reference region.

[0091] In some embodiments, the first training image should be acquired before the second training image is acquired. This is because the first training image should represent similar image conditions to the runtime images obtained during the production phase. As is known, during image acquisition by an e-beam tool, the surface of the sample is scanned by a focused e-beam, and the sample constantly collects electrical charges. The accumulation of surface charges caused by the e-beam on the sample can lead to image artifacts such as, for example, image distortion, changes related to gray scale, contrast, edge sharpness, etc. Therefore, in order to acquire the first training image in similar conditions to the runtime images, for each given region, the first training image should be acquired first, before further charging effects accumulate on the sample, similar to when the runtime images are acquired at runtime. The second image can be acquired after the first image is acquired, without having to change any tool configuration.

[0092] In some embodiments, the first training image is acquired first, for example by acquiring N1 frames. Then, the second training image is acquired, for example by acquiring N2 frames (N2 > N1). In some cases, the N1 frames can be a portion of the N2 frames. For example, the SEM can first acquire N1 frames, which will be combined into the first training image. Then, the SEM can continue to acquire (N2 - N1) frames, which will be combined with the N1 frames to form the second training image. In some other cases, the N2 frames can be acquired separately, which do not include the N1 frames.

[0093] In some embodiments, optionally, the training images can include multiple channels captured from different perspectives. For example, one channel can be taken from a top detector of the inspection tool from a vertical perspective, while another channel can be taken from a side detector of the inspection tool from a side perspective. In some cases, there can be more than one side detector from different angles, and thus, the training images can include multiple side channel images. In some cases, the multiple channel images can be combined into one combined image.

[0094] The second training images (e.g., by the training set generator 104) having high SNR are then labeled (406). As previously noted, it is preferable to perform the labeling on the high SNR images in order to improve the accuracy of the labeled label data that will be used as ground truth data for training the ML model. The label data can be obtained in various ways. For example, the label data can be obtained by manual labeling, or can be synthetically generated (e.g., using design data, such as CAD-based images). As another example, the ground truth data can be generated based on machine learning. For example, a machine learning model can be trained using manually labeled images, and the trained model can be used to automatically (or semi-automatically) generate label data for input images. An example of a machine learning based label data generation system is described in U.S. Patent Application No. 16 / 942,677, entitled GENERATING TRAINING DATA USABLE FOR EXAMINATION OF A SEMICONDUCTOR SPECIMEN, which is incorporated by reference herein in its entirety. In some cases, the label data can be generated as a combination of any of the above. The present disclosure is not limited to the particular manner in which the label data for the second training images is obtained.

[0095] The first and second training images can be registered (408) to correct for a shift between them, so that the first training image and the label data are aligned. The shift between the two training images can be caused by various factors, such as, for example, drift caused by charging effects, drift caused by the working point of the tool (e.g., scanner and / or stage drift), and / or drift caused by shrinkage of the sample, etc. Image registration as referred to in the present disclosure can include measuring the shift between the two images, and shifting one image relative to the other to correct for the shift. In particular, in the present disclosure, once the shift is measured, the labeled label data in the second training image (i.e., the high SNR image) can be shifted (along with or without the second image itself) according to the shift (relative to the first training image) in order to align with the low SNR image. Alternatively, the first training image can be shifted according to the shift (relative to the second training image) to align with the label data (and with the second training image).

[0096] Registration can be performed using any suitable registration algorithm known in the art. For example, registration can be performed using one or more of the following algorithms: region-based algorithms, feature-based registration, or phase-correlated registration. An example of a region-based method is registration using optical flow, such as the Lucas-Cannard (LK) algorithm. Feature-based methods are based on finding different information points (“features”) in the two images and calculating the desired transformation between each pair according to the correspondence of features. This allows for flexible registration (i.e., non-rigid registration), where different regions are moved individually. Phase-correlated registration (PCR) is performed using frequency domain analysis (where the phase difference in the Fourier domain is converted into registration in the image domain).

[0097] Therefore, training samples for training the ML model are generated (410), including registered first training images and label data. In some embodiments, one or more additional training samples may be obtained from one or more additional reference regions in a similar manner.

[0098] Including, for example Figure 4 The training set of one or more generated training samples can be used to train the ML model in a supervised manner. After training, the ML model can be used to process runtime images representing inspection regions of design patterns that share the same design pattern as the samples and obtain inspection data specific to a given inspection application. The runtime images are acquired with a relatively low SNR (same / similar to the low SNR of the first training image as described above).

[0099] Figure 5 A schematic diagram illustrating the training process of an ML model according to certain embodiments of the currently disclosed subject matter is shown. During the training phase, data can be obtained according to... Figure 2 and Figure 4application. Specifically, a first training image 502 is acquired that is a low SNR image representing a reference region of a specimen. A second training image 504 is acquired that is a high SNR image of the same reference region. The second training image 504 is labeled to obtain its label data (506) (e.g., a segmentation map associated with the second training image in the example of a segmentation application). The two images 502 and 504 are registered so that either the label data or the first training image 502 is fixed according to the offset between them. From this, a training sample is generated that includes the registered first training image 502 and the label data 506. The training sample can be used to train (508) a ML model 510, thereby obtaining a trained ML model characterized by segmentation-related training parameters. In some embodiments, the training process can be iterative and can be repeated several times until the ML model is sufficiently trained, e.g., is able to provide output of a segmentation map that meets accuracy criteria. For example, a cost function related to segmentation accuracy (e.g., ground truth label data versus predicted segmentation map) can be used to train the ML model.

[0100] Although Figure 5 Although only one training sample is shown in FIG. 5, this is not intended to limit the disclosure in any way. In some embodiments, one or more additional training samples can be obtained in a similar manner, and the training process can be repeated using the additional training samples. In some cases, optionally, a validation image set can be used to validate the trained ML model. The image validation set can be a different set of images than the training set, and can include images selected for validation purposes. A user can provide feedback on the results achieved by the ML model during training or validation.

[0101] Referring now to Figure 3 FIG. 6 shows a generalized flowchart of runtime inspection of a semiconductor specimen using a trained ML, in accordance with certain embodiments of the presently disclosed subject matter.

[0102] During runtime inspection of a specimen, a runtime image (e.g., an FP image as described above) representing an inspection region of a semiconductor specimen can be obtained (302) (e.g., by inspection tool 120). The runtime image has a relatively low signal-to-noise ratio (SNR) (e.g., relative to a high SNR as described herein, or relative to a threshold). For example, the runtime image can be acquired by an e-beam tool such as, for example, a SEM.

[0103] The runtime image can be processed (304) using a machine learning (ML) model (e.g., ML model 108 in PMC 102) to obtain inspection data specific to a given inspection application. The ML model can have been previously trained for the given inspection application using one or more training samples, each representing a respective reference region sharing the same design pattern as the inspection region, as described above with respect to FIG. 5. The ML model can be configured to generate inspection data specific to the given inspection application based on the input runtime image. For example, the ML model can be configured to generate inspection data specific to a given inspection application based on the input runtime image and the training parameters of the trained ML model.Figure 2 and Figure 4 Specifically, each training sample includes a first training image (204) with a relatively low signal-to-noise ratio (SNR) (e.g., relatively lower than a high SNR or threshold) for the corresponding reference region and label data (206) indicating the ground truth for a given inspection application in the corresponding reference region. The label data is obtained by labeling a second training image with a relatively high SNR for the corresponding reference region. As mentioned above, the first training image is acquired under similar imaging conditions (e.g., electron dose, number of frames, etc.) to the runtime image, such that the low SNR of the first training image is the same / similar (e.g., identical, similar, or highly correlated) as the low SNR of the runtime image. Therefore, the low SNR of the first training image and the low SNR of the runtime image can be similar / equivalent to some extent, but not necessarily identical. Various similarity measures and algorithms can be used to determine the degree / level of equivalence / similarity between them.

[0104] As previously mentioned, the inspection applications mentioned herein can be any application including (but not limited to) the following: a segmentation application for segmenting a runtime image into one or more segments in the inspection region; a metrology application for obtaining one or more measurements of structural features in the inspection region; a defect detection application for detecting one or more candidate defects in the inspection region; and a defect classification application for classifying one or more candidate defects in the inspection region.

[0105] According to some embodiments, the inspection application may be a segmentation application. As used herein, the term "segmentation" can refer to any process that divides an image into meaningful parts / fragments (e.g., background and foreground, noisy and non-noisy regions, various structural elements, defects and non-defects, etc.) while providing per-pixel or per-region values ​​indicative of such fragments. In such cases, training samples may include a first training image (e.g., an SEM image) and label data, which may be a ground truth segmentation map corresponding to the SEM image (e.g., indicating a specific fragment to which each pixel of at least a portion of the image in the runtime belongs). For example, in some cases, a fragment may correspond to one or more structural elements presented in the first training image. As used herein, a structural element can refer to any primitive object on image data having a geometric shape or structure with an outline, in some cases combined with other objects. Structural elements may be presented, for example, in the form of polygons.

[0106] After training the ML model, the trained ML model can be used to process runtime images and output predicted segmentation maps, which provide information about predicted labels associated with corresponding pixels in the runtime image. Each predicted label indicates the segment in the runtime image to which the corresponding pixel belongs.

[0107] Figure 6 An example of a first training image according to certain embodiments of the presently disclosed subject matter is shown. Training image 602 is illustrated as an SEM image captured by a SEM tool and representing a region of a die of a wafer. As shown, there are multiple structure elements 604 (shown as polygons representing elements of contacts on the wafer) in the image. In this example, label data (obtained from annotation on a second training image) can be provided as a segmentation map, such as a binary map representing two segments, a first segment corresponding to the structure elements in the image, and a second segment corresponding to the background region.

[0108] Figure 7 is a schematic diagram of a runtime inspection process using a ML model according to certain embodiments of the presently disclosed subject matter. During runtime, a runtime image 702 of a sample with low SNR is acquired and processed by the trained ML model 704 to obtain inspection data specific to a given inspection application. For example, in a segmentation application, the output inspection data can be a segmentation map 706 corresponding to the runtime image 702. The obtained segmentation map can provide information indicating per-pixel or per-region segmentation labels of different segments on the image. For example, polygons on one layer can have one segmentation label, and polygons on another layer can have a different segmentation label, while the background can have a separate segmentation label.

[0109] In some embodiments, such a segmentation map can be used by metrology tools to perform measurements on the sample. As another example, it can also be used in building attributes, e.g., for defining whether a defect is on the main pattern, on the background, or both, or can be used in ADR for applying segment-specific detection thresholds on each segment, etc.

[0110] In some embodiments, the inspection application is a defect classification application. The label data acquired in this case can provide information of the class of defects (e.g., particle, pattern deformation, bridge, etc.) present in the reference region and optionally the probability that the defects belong to these classes. The cost function used during ML training can be based on classification error between the predicted class and its true value (class label).

[0111] In some embodiments, the inspection application is a defect detection application. The label data acquired in this case can provide information of whether a candidate defect in a list of candidate defects appearing in the reference region is a defect of interest (DOI) or nuisance. For example, the label data can be provided, e.g., in the form of DOI bounding boxes, or in the form of a binary image where only pixels belonging to DOIs obtain a value of “1”, while non-defect pixels obtain a value of “0”, etc. The cost function used during ML training can be based on detection accuracy / capture rate, and optionally, also on penalties for false detection and over-detection.

[0112] In some embodiments, the inspection tool is a metrology application. The label data acquired in this case can provide information about one or more measured values (e.g., CD measurements) of the structural elements in the inspection region. The cost function used during training of the ML can be based on the measurement accuracy of the predicted measurements relative to the true measurements.

[0113] According to some embodiments, the ML model can be implemented as a classifier. The terms “classifier”, “classifier model” or “classification model” mentioned herein should be interpreted broadly to encompass any learning model capable of identifying, based on a training set, to which of a set of classes / classes a new instance belongs. For example, in the exemplified segmentation application, a classifier can be trained to classify candidate pixels into a set of segmentation classes as defined by a user. The trained classifier can be used for image segmentation, i.e., for providing, for each pixel in an image, a predicted label indicative of the segment it belongs to. It is noted that the classifier can be implemented as various types of machine learning models such as, for example, linear classifiers, support vector machines (SVMs), neural networks, decision trees, etc., and the present disclosure is not limited to a particular model implemented with the present disclosure.

[0114] According to certain embodiments, the training process as described above with reference to Figure 2 and Figure 4 may be included as part of the process for generating an inspection recipe that can be used by the system 101 and / or the inspection tool 120 for run-time online inspection (where the ML model, once trained, can be used as part of the inspection recipe). Thus, the presently disclosed subject matter also includes systems and methods for generating an inspection recipe during the recipe setup phase, as described with reference to Figure 2 and Figure 4 (and various embodiments thereof). It is noted that the term “inspection recipe” should be interpreted broadly to encompass any recipe that can be used by an inspection tool for any inspection application including the embodiments as described above.

[0115] It is noted that the examples shown in the present disclosure, such as, for example, the various ways of obtaining the first and second training images, the exemplified inspection applications and label data, etc., are shown for exemplary purposes and should not be considered as limiting the present disclosure in any way. Other suitable examples / implementation can be used in addition to or instead of the above.

[0116] One of the advantages of certain embodiments of the training process as described herein is that it enables acquisition of labels from high SNR images and associating them with low SNR images, where the label data and low SNR images are used together as training samples for training the ML model. This ensures the accuracy of the acquired label data, which are used as ground truth for training, while enabling the ML model to be trained on low SNR images that are similar in imaging conditions to the runtime images that will be checked at runtime, thereby improving the performance of the trained ML model in runtime checks in terms of robustness and accuracy.

[0117] A further advantage of certain embodiments of the training process as described herein is that it enables inspection of certain sensitive layers of a sample that were previously not possible to be inspected, as these layers require minimal electron dose during inspection in order to reduce pattern damage, which results in very low SNR of the images, and annotation on such low SNR images cannot be performed.

[0118] A further advantage of certain embodiments of the training process as described herein is that since annotation on high SNR images is easier and faster, this further improves the user experience of annotation and the time to recipe (TTR) for training the ML.

[0119] It will be understood that the disclosure is not limited in its application to the details set forth in the description herein contained or exemplified in the drawings.

[0120] It will also be understood that a system according to the present disclosure can be implemented at least partially in suitable programmed computers. Likewise, the present disclosure contemplates a computer program being readable by a computer useable to perform a method according to the steps of the method of the present disclosure. The present disclosure further contemplates a non-transitory computer-readable storage medium having stored thereon, computer-executable instructions that when implemented by a computer cause the computer to carry out a method according to the steps of the method of the present disclosure.

[0121] The present disclosure is capable of other embodiments and of being practiced or being carried out in various ways. Accordingly, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of "including," "comprising," "having" and "containing" and variations thereof herein is intended to be equivalent to the term "consisting of" to mean the inclusion, but not the exclusion, of something. Thus, these terms and variations thereof are used in the following claims to enable open and unrestricted use of the claims as being allied in all possible ways with the other elements.

[0122] Those skilled in the art will readily recognize various modifications and changes that can be made to the embodiments of the present disclosure as described above without departing from the scope of the present disclosure, which is defined by the following claims and their equivalents.

Claims

1. A computerized system of runtime inspection of a semiconductor sample, the system comprising processing and memory circuitry (PMC) configured to: obtain a runtime image representing an inspection area of the semiconductor sample, the runtime image having a low signal-to-noise ratio (SNR); and process the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model is previously trained for the given inspection application using one or more training samples, each training sample representing a respective reference area sharing a same design pattern as the inspection area and comprising: a first training image of the respective reference area, the first training image having a low SNR similar to the low SNR of the runtime image; and label data indicating ground truth about the given inspection application in the respective reference area, the label data obtained by annotating a second training image of the respective reference area having a high SNR, wherein the high SNR is defined to ensure a level of annotation accuracy on the second training image while maintaining a correspondence between the second training image and the first training image.

2. The computerized system of claim 1, wherein the runtime image and the one or more training samples are acquired by an e-beam tool.

3. The computerized system of claim 1, wherein the given inspection application is one of: a segmentation application for segmenting the runtime image into one or more segments in the inspection area, a metrology application for obtaining one or more measurement values about structural elements in the inspection area, a defect detection application for detecting one or more candidate defects in the inspection area, and a defect classification application for classifying one or more defects in the inspection area.

4. The computerized system of claim 1, wherein the first training image is generated based on a first number of training frames acquired for the reference area, and the second training image is generated based on a second number of training frames acquired for the reference area, wherein the first number is less than the second number, and wherein the runtime image is generated based on a first number of runtime frames.

5. The computerized system of claim 4, wherein the second number of training frames includes the first number of training frames.

6. The computerized system of claim 1, wherein the first training image is generated based on a first dose of electrons, and the second training image is generated based on a second dose of electrons, wherein the first dose is less than the second dose, and wherein the runtime image is generated based on the first dose of electrons.

7. The computerized system of claim 1, wherein the label data is obtained based on at least one of: manual annotation, synthetically generated labels based on design data, machine learning derived labels, or a combination thereof.

8. The computerized system of claim 1, wherein the first training image and the second training image are registered to correct for a shift between them, such that the first training image and the label data are aligned.

9. The computerized system of claim 1, wherein the first training image is acquired prior to acquiring the second training image.

10. The computerized system of claim 1, wherein the inspection region is a portion of the semiconductor sample that can only be inspected via images having a low SNR.

11. The computerized system of claim 1, wherein the given inspection application is a segmentation application for segmenting the runtime image into one or more segments in the inspection region, and the label data indicates a particular segment to which each pixel of at least a portion of the runtime image in the one or more segments belongs.

12. The computerized system of claim 1, wherein the inspection region is from an inspection die of the semiconductor sample, and the respective reference region is from a reference die of the inspection die, the reference die being from the semiconductor sample or from a different semiconductor sample.

13. A computerized method of training a machine learning (ML) model that can be used for inspecting a semiconductor sample, the method performed by processing and memory circuitry (PMC) and comprising: obtaining a training set, the training set comprising one or more training samples, each training sample representing a respective reference region having a given design pattern, the training sample comprising: a first training image, the first training image having a low signal-to-noise ratio (SNR); and label data, the label data indicating a ground truth about a given inspection application in the respective reference region, the label data obtained by annotating a second training image of the respective reference region having a high SNR, wherein the high SNR is defined to ensure a level of annotation accuracy on the second training image while maintaining a correspondence between the second training image and the first training image; and training the ML model using the training set for the given inspection application; wherein the ML model, when trained, can be used to process a runtime image representing an inspection region of the semiconductor sample that shares a same design pattern as the given design pattern and obtain inspection data specific to the given inspection application, the runtime image having a low SNR similar to the low SNR of the first training image.

14. The computerized method of claim 13, wherein the given inspection application is one of: a segmentation application for segmenting the runtime image into one or more segments in the inspection region, a metrology application for obtaining one or more measurements about a structural element in the inspection region, a defect detection application for detecting one or more candidate defects in the inspection region, and a defect classification application for classifying one or more defects in the inspection region.

15. The computerized method of claim 13, wherein said obtaining the training set comprises generating the first training image based on a first number of training frames acquired for the reference region, and generating the second training image based on a second number of training frames acquired for the reference region, wherein the first number is less than the second number, and wherein the runtime image is generated based on the first number of runtime frames.

16. The computerized method of claim 13, wherein said obtaining the training set comprises generating the first training image based on a first dose of electrons, and generating the second training image based on a second dose of electrons, wherein the first dose is less than the second dose, and wherein the runtime image is generated based on the first dose of electrons.

17. The computerized method of claim 13, wherein said obtaining the training set comprises registering the first training image and the second training image to correct for a shift between them, such that the first training image and the label data are aligned.

18. The computerized method of claim 13, wherein the first training image is acquired prior to acquiring the second training image.

19. The computerized method of claim 13, wherein the given inspection application is a segmentation application for segmenting the runtime image into one or more segments in the inspection region, and the label data indicates a particular segment to which each pixel of at least a portion of the runtime image in the one or more segments belongs.

20. A non-transitory computer-readable storage medium tangibly embodying a program of instructions, the program of instructions, when executed by a computer, causing the computer to perform a method of runtime inspection of a semiconductor sample, the method comprising: obtaining a runtime image representing an inspection region of the semiconductor sample, the runtime image having a low signal-to-noise ratio (SNR); and processing the runtime image using a machine learning (ML) model to obtain inspection data specific to a given inspection application, wherein the ML model was previously trained for the given inspection application using one or more training samples, each training sample representing a respective reference region sharing a same design pattern as the inspection region and comprising: a first training image of the respective reference region having a low SNR similar to the low SNR of the runtime image; and label data indicating ground truth in the respective reference region pertaining to the given inspection application, the label data obtained by annotating a second training image of the respective reference region having a high SNR, wherein the high SNR is defined to ensure a level of annotation accuracy on the second training image while maintaining a correspondence between the second training image and the first training image.

Images