Defect Classification in Semiconductor Samples Based on Machine Learning

Abstract

A method and system for automatic defect classification are provided. The method includes: obtaining data providing information of a set of physical properties of defects, wherein the physical properties of the set of defects can be used to distinguish defects of different categories in a plurality of categories; training a first machine learning model to generate a multi-label output vector of information providing values ​​of the physical properties for a given defect, thereby generating a multi-label descriptor for the given defect; and using the trained first machine learning model to generate a multi-label descriptor of a defect in a sample. The method may further include: obtaining data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category in a plurality of categories and including a unique set of values ​​of the physical properties; and classifying the defects in the sample by matching the multi-label descriptors of the defects generated respectively with the multi-label dataset.

Description

Technical Field

[0001] The presently disclosed subject matter relates generally to the field of specimen inspection, and more particularly to automating defect classification in specimens. Background Art

[0002] Current demands for high density and performance associated with ultra-large scale integration of manufactured devices require sub-micron features, increased transistor and circuit speeds, and improved reliability. Such demands require forming device features with high precision and uniformity, which in turn requires careful monitoring of the manufacturing process, including automated inspection of the device while it is still in semiconductor wafer form.

[0003] The term "sample" as used in this specification should be broadly interpreted to cover any type of wafer, mask and other structure, combination and / or portion thereof used to manufacture semiconductor integrated circuits, magnetic heads, flat panel displays and other semiconductor manufactured articles.

[0004] The term "inspection" used in this specification should be broadly interpreted as covering any kind of metrology-related operations and operations related to the detection and / or classification of defects in a sample during sample manufacturing. Inspection is provided by using a non-destructive inspection tool during or after the manufacture of the sample to be inspected. As a non-limiting example, the inspection process may include runtime scanning (in single or multiple scans), sampling, reviewing, measuring, classification and / or other operations provided by the same or different inspection tools for a sample or a portion of a sample. Similarly, inspection may be provided before the manufacture of the sample to be inspected, and the inspection may include, for example, generating (multiple) inspection recipes and / or other setup operations. It should be noted that, unless otherwise specifically stated, the term "inspection" or its derivatives used in this specification have no restrictions on the resolution or size of the inspection area. As non-limiting examples, various non-destructive inspection tools include scanning electron microscopes, atomic force microscopes, optical inspection tools, and the like.

[0005] As a non-limiting example, the runtime inspection may employ a two-stage process, e.g., after inspecting a sample, the sample locations are then reviewed for potential defects. During the first stage, the surface of the sample is inspected at high speed and relatively low resolution. In the first stage, a defect map is generated to show suspect locations on the sample with a high probability of defects. During the second stage, at least a portion of the suspect locations are more thoroughly analyzed at a relatively high resolution. In some cases, the two stages may be implemented by the same inspection tool, while in other cases, the two stages may be implemented by different inspection tools.

[0006] Inspection processes are used at various steps during semiconductor manufacturing to detect and classify defects on samples. The efficiency of inspection can be increased by automating the process(es), for example, Automatic Defect Classification (ADC), Automatic Defect Review (ADR), etc. Summary of the invention

[0007] According to certain aspects of the presently disclosed subject matter, a method for automatically classifying defects in a semiconductor sample into a plurality of categories is provided. The method is performed by a processing and memory circuit system (PMC) and includes: obtaining data providing information of a set of physical properties of defects, the physical properties of the set of defects being usable to distinguish defects of different categories in a plurality of categories; and after training a first machine learning model to process a sample including one or more images providing information of a given defect so as to generate a multi-label output vector providing information of a value of a physical property from a set of physical properties for the given defect, thereby generating a multi-label descriptor for the given defect, using the trained first machine learning model to generate a multi-label descriptor of the defect in the sample, the descriptor being usable for classification.

[0008] The method may further include: obtaining data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category in a plurality of categories and including a unique set of values ​​of a physical property from a set of physical properties; and after training a second machine learning model to provide multi-label classification, classifying defects in the sample using the trained second machine learning model by matching the respectively generated multi-label descriptors of the defects with the multi-label dataset. Alternatively or additionally, the generated multi-label descriptors of the defects in the sample may be analyzed to identify multimodal behavior of one or more categories and / or identify new repeated multi-label datasets, thereby identifying new categories of defects.

[0009] According to other aspects of the presently disclosed subject matter, a system for classifying defects in a semiconductor sample into multiple categories is provided. The system includes a processing and memory circuit system (PMC) operably connected to an input interface, wherein the input interface is configured to receive a sample including an image providing information of a defect; and wherein the PMC is configured to: obtain data providing information of a set of physical properties of defects, wherein the set of physical properties of defects can be used to distinguish defects of different categories in a plurality of categories; and after training a first machine learning model to process a sample including one or more images providing information of a given defect so as to generate a multi-label output vector providing information of values ​​of physical properties from a set of physical properties for the given defect, thereby generating a multi-label descriptor for the given defect, use the trained first machine learning model to generate a multi-label descriptor of the defect in the sample, wherein the descriptor can be used for classification.

[0010] The PMC can be further configured to: obtain data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category among multiple categories and including a unique set of values ​​of physical properties from a set of physical properties; and after training a second machine learning model to provide multi-label classification, use the trained second machine learning model to classify defects in the sample by matching the multi-label descriptors of the defects generated respectively with the multi-label dataset.

[0011] The PMC can be further configured to: obtain data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category among multiple categories and including a unique set of values ​​of a physical property from a set of physical properties; and analyze the generated multi-label descriptors of defects in the sample to identify a new repeated multi-label dataset, thereby identifying a new category of defects and / or identify multimodal behavior of one or more categories by analyzing the generated multi-label descriptors of defects in the sample and / or identify a new repeated multi-label dataset, thereby identifying the new category of the defect.

[0012] According to a further aspect, and optionally in combination with other aspects of the presently disclosed subject matter, classifying defects may include: defining a certainty threshold as a ratio between the number of values ​​in the respectively generated multi-label descriptors that match a given category and the total number of values ​​in the multi-label dataset indicating the given category. As non-limiting examples, the certainty threshold may enable at least one of: optimizing the confidence level of defect classification; identifying misclassified defects; setting purity requirements for each category and / or group of categories separately; setting accuracy requirements for each category and / or group of categories separately; setting extraction requirements for each category and / or group of categories separately.

[0013] According to further aspects, and optionally in combination with other aspects of the presently disclosed subject matter, a set of physical properties of a defect can provide information on at least one of: physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, layer, texture, edge, and chemical composition.

[0014] As a non-limiting example, the multiple categories may include a "particle" category and a "bridge" category, where the physical properties of a corresponding set of defects may include the roughness of the texture, the sharpness of the edge, the position relative to the top of the pattern, and the position relative to the two patterns.

[0015] According to a further aspect, and optionally, in combination with other aspects of the presently disclosed subject matter, the values ​​of the physical attributes in the multi-label descriptors and multi-label datasets can be binary. Alternatively, the values ​​of the physical attributes in the multi-label descriptors and multi-label datasets can correspond to "yes", "no" and "irrelevant" about the corresponding physical attributes.

[0016] An advantage of certain embodiments of the presently disclosed subject matter is the ability to classify defects using learned attributes that have physical meaning, thereby enabling physical understanding and debugging, while having inherent high performance capabilities for the learned attributes.

[0017] A further advantage of certain embodiments of the presently disclosed subject matter is the ability to classify into unseen categories by defining such categories through physically meaningful attributes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to understand the invention and to see how it may be carried out in practice, embodiments will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which:

[0019] Figure 1 shows a functional block diagram of an inspection system according to certain embodiments of the presently disclosed subject matter;

[0020] Figure 2 A generalized flow chart illustrating machine learning based defect classification according to certain embodiments of the presently disclosed subject matter; and

[0021] Figure 3 A generalized flow chart illustrating setup steps for classifying defects according to certain embodiments of the presently disclosed subject matter. DETAILED DESCRIPTION

[0022] In the following detailed description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be appreciated by those skilled in the art that the subject matter disclosed herein may be practiced without these specific details. In other cases, well-known methods, processes, components, and circuits are not described in detail to avoid obscuring the subject matter disclosed herein.

[0023] Unless otherwise specifically stated, as will be apparent from the following discussion, it should be understood that throughout the specification discussion, use of terms such as "processing," "computing," "representing," "comparing," "generating," "training," or similar terms refers to computer actions and / or processes that manipulate and / or transform data into other data, the data being represented as physical (such as electronic) quantities and / or the data representing physical objects. The term "computer" should be broadly interpreted to cover any kind of hardware-based electronic device with data processing capabilities, including, as a non-limiting example, the FPEI (Manufacturing Process Inspection Information) system disclosed in this application and its corresponding parts.

[0024] The terms "non-transitory memory" and "non-transitory storage medium" as used herein should be broadly interpreted to cover any volatile or non-volatile computer memory suitable for the presently disclosed subject matter.

[0025] The term "defect" as used in this specification should be interpreted broadly to cover any kind of abnormal or undesirable feature formed on or in a sample.

[0026] The term "design data" as used in this specification should be broadly interpreted to cover any data indicative of a hierarchical physical design (layout) of a sample. The design data may be provided by a corresponding designer and / or may be derived from the physical design (e.g., through complex simulations, simple geometric and Boolean operations, etc.). The design data may be provided in different formats, such as, as non-limiting examples, GDSII format, OASIS format, etc. The design data may be presented in vector format, grayscale intensity image format, or other manners.

[0027] It should be understood that, unless otherwise specifically stated, certain features of the currently disclosed subject matter described in the context of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the currently disclosed subject matter described in the context of a single embodiment may also be provided individually or in any suitable sub-combination. In the following detailed description, many specific details are set forth in order to provide a thorough understanding of the method and apparatus.

[0028] With this in mind, turn your attention to Figure 1 , Figure 1 A functional block diagram of an inspection system according to certain embodiments of the presently disclosed subject matter is shown. Figure 1The inspection system 100 shown in the figure can be used for inspection of samples (e.g., samples of wafers and / or portions thereof) as part of a sample manufacturing process. The inspection system 100 shown includes a computer-based system 103 that is capable of automatically determining information related to metrology and / or defects using images obtained during sample manufacturing. Such images are referred to as manufacturing process (FP) images hereinafter. The system 103 is referred to as the FPEI (manufacturing process inspection information) system hereinafter. The FPEI system 103 can be operably connected to one or more low-resolution inspection tools 101 and / or one or more high-resolution inspection tools 102 and / or other inspection tools. The inspection tool is configured to capture FP images and / or review the captured (multiple) FP images and / or enable or provide measurements related to the captured (multiple) images. The FPEI system can be further operably connected to a CAD server 110 and a data repository 109.

[0029] The FPEI system 103 includes a processor and memory circuit system (PMC) 104, which is operably connected to a hardware-based input interface 105 and to a hardware-based output interface 106. The PMC 104 is configured to provide all processing required for operating the FPEI system, as described with reference to FIG. Figure 2-3 As further described in detail, the PMC 104 includes a processor (not shown separately) and a memory (not shown separately). The processor of the PMC 104 can be configured to execute several functional modules according to computer-readable instructions implemented on a non-transitory computer-readable memory included in the PMC. Such functional modules are hereinafter referred to as being included in the PMC. The functional modules included in the PMC 104 include a machine learning module 111 for the generation of descriptors (hereinafter referred to as a generator of descriptors) and a machine learning module 112 for the classification of defects (hereinafter referred to as a classifier) ​​that are operably connected.

[0030] The operation of the FPEI system 103, PMC 104 and the functional blocks therein will be referred to as Figure 2-3 Described in further detail.

[0031] If you will refer to Figure 2-3As further described in detail, the FPEI system is configured to receive FP input data via an input interface 105. The FP input data may include data generated by an inspection tool (and / or its derivatives and / or metadata associated therewith) and / or data generated and / or stored in one or more data repositories 109 and / or a CAD server 110 and / or other related data repositories. It should be noted that the FP input data may include images (e.g., captured images, images derived from captured images, simulated images, synthetic images, etc.) and associated numerical data (e.g., metadata, manually constructed attributes, etc.). It should also be noted that the image data may include data related to a layer of interest and / or to one or more other layers of the sample. Optionally, for training purposes, the FP input data may include the entire available FAB data or portions thereof selected according to certain criteria.

[0032] The FPEI system is further configured to process at least a portion of the received FP input data and send the results (or portions thereof) to a storage system 107, to an inspection tool(s), to a computer-based graphical user interface (GUI) 108 for presenting the results, and / or to an external system (e.g., a yield management system (YMS) of the FAB) via an output interface 106. The GUI 108 may be further configured to enable user-specified input related to operating the FPEI system 103.

[0033] As a non-limiting example, the sample may be inspected by one or more low-resolution inspection machines 101 (e.g., an optical inspection system, a low-resolution SEM, etc.). The resulting data providing information of a low-resolution image of the sample (hereinafter referred to as low-resolution image data 121) may be transmitted directly to the FPEI system 103 or transmitted to the FPEI system 103 via one or more intermediate systems. Alternatively or additionally, the sample may be inspected by a high-resolution machine 102 (e.g., a subset of potential defect locations selected for review may be reviewed by a scanning electron microscope (SEM) or an atomic force microscope (AFM). The resulting data providing information of a high-resolution image of the sample (hereinafter referred to as high-resolution image data 122) may be transmitted directly to the FPEI system 103 or transmitted to the FPEI system 103 via one or more intermediate systems.

[0034] It should be noted that images of desired locations on the sample may be captured at different resolutions. As a non-limiting example, so-called "defect images" of desired locations may be used to distinguish defects from false alarms, while so-called "class images" of desired locations are obtained at a higher resolution and may be used for defect classification. In some embodiments, images of the same location (with the same or different resolutions) may include several images registered therebetween (e.g., an image captured from a given location and one or more reference images corresponding to the given location).

[0035] When processing FP input data (e.g., low-resolution image data and / or high-resolution images, optionally together with other data (e.g., design data, synthesis data, etc.)), the FPEI system can send results (e.g., instruction-related data 123 and / or 124) to any of the inspection tools (multiple inspection tools), store the results (e.g., defect attributes, defect classification, etc.) in the storage system 107, present the results via the GUI 108, and / or send the results to an external system (e.g., to a YMS).

[0036] Those skilled in the art will readily appreciate that the teachings of the presently disclosed subject matter are not limited to Figure 1 limitations of the systems shown; equivalent and / or modified functionality may be combined or divided in another manner and implemented in any suitable combination of software and firmware and / or hardware.

[0037] Without limiting the scope of the present disclosure in any way, it should also be noted that the inspection tool can be implemented as various types of inspection machines, such as optical imaging machines, electron beam inspection machines, etc. In some cases, the same inspection tool can provide low-resolution image data and high-resolution image data. In some cases, at least one inspection tool can have metrology capabilities.

[0038] The descriptor generator 111 and the classifier 112 can be implemented as separate or combined machine learning modules. For illustrative purposes only, the following description is provided for a machine learning module (descriptor generator 111 and classifier 112) implemented as a deep neural network (DNN). It will be readily appreciated by those skilled in the art that the teachings of the subject matter currently disclosed are equally applicable to other suitable techniques based on machine learning.

[0039] The descriptor generator 111 and the classifier 112 may include one or more DNN subnets, each of which includes multiple layers organized according to a corresponding DNN architecture. Optionally, at least one of the DNN networks may have an architecture different from that of the other DNN networks. As a non-limiting example, the layers in the corresponding DNN network may be organized according to a convolutional neural network (CNN) architecture, a recurrent neural network architecture, a recursive neural network architecture, or in other ways. Optionally, at least some of the DNN subnets may have one or more common layers (e.g., a final fusion layer, an output fully connected layer, etc.). The output of the descriptor generator 111 may be used as an input to the classifier 112.

[0040] Each layer of the DNN network may include multiple basic computing elements (CEs), which are generally referred to as dimensions, neurons, or nodes in the art. The computing elements of a given layer may be connected to the CEs of the previous layer and / or the next layer. Each connection between the CE of the previous layer and the CE of the next layer is associated with a weighted value. A given CE may receive input from the CE of the previous layer via a corresponding connection, and each given connection is associated with a weighted value, which may be applied to the input of a given connection. The weighted value may determine the relative strength of the connection, and thereby determine the relative influence of the corresponding input on the output of a given CE. A given CE may be configured to calculate an activation value (e.g., a weighted sum of inputs), and further derive an output by applying an activation function to the calculated activation. The activation function may be, for example, an identity function, a deterministic function (e.g., linear, S-shaped, threshold, etc.), a random function, or other suitable function. The output may be transmitted from a given CE to a next layer of CE via a corresponding connection. Similarly, as described above, each connection at the output of a CE may be associated with a weighted value, which may be applied to the output of the CE before being received as an input to the CE of the next layer. In addition to weighted values, there may also be thresholds (including limiting functions) associated with connections and CEs.

[0041] Before training, the weights and / or thresholds of the deep neural network can be first selected, and the weights and / or thresholds can be further iteratively adjusted or modified during training to achieve a set of optimal weights and / or thresholds in the trained DNN module. After each iteration, the difference between the actual output generated by the DNN module and the target output associated with the corresponding training data set can be determined. The difference can be referred to as an error value. When the cost function indicating the error value is less than a predetermined value, or when a limited change in performance is achieved between iterations, it can be determined that the training is complete. The generator 111 of the descriptor can be trained separately from the classifier 112. In the following, a set of input data used to train the corresponding machine learning model is referred to as a training set. For DNN, the training set is used to adjust the weights / thresholds of the deep neural network.

[0042] It should be noted that the teachings of the presently disclosed subject matter are not limited by the architecture (including the number and / or architecture of the DNN networks) of the descriptor generator 111 and the classifier 112. As a non-limiting example, the classifier 112 can operate in the manner disclosed in PCT application PCT / IL2019 / 050155 filed on February 7, 2019, which is incorporated herein by reference in its entirety.

[0043] It should be noted that Figure 1 The inspection system shown can be implemented in a distributed computing environment, where Figure 1 The above-mentioned functional modules shown can be distributed on several local and / or remote devices and can be linked through a communication network. It should also be noted that in other embodiments, at least part of the inspection tools 101 and / or 102, the data repository 109, the storage system 107 and / or the GUI 108 can be external to the inspection system 100 and operate in a manner of data communication with the FPEI system 103 via the input interface 105 and the output interface 106. The FPEI system 103 can be implemented as (multiple) independent computers used in conjunction with the inspection tool. Alternatively, the corresponding functions of the FPEI system can be at least partially integrated with one or more inspection tools.

[0044] Reference Figure 2 , shows a general flow chart of machine learning based classification according to some embodiments of the presently disclosed subject matter. The process includes a setup step (201), the setup step (201) includes training (202) a descriptor generator 111 to provide class-dependent descriptors for defects in corresponding images, and training (203) a classifier 112 to provide a multi-label classification of defects based on the corresponding descriptors. The setup step will refer to Figure 3 Further details are given.

[0045] During runtime (204), the PMC of the FPEI system processes (205) a FP sample including one or more FP images using the obtained trained descriptor generator 111 and classifier 112. Thereby, the PMC obtains (206) classification-related data characterizing at least one of the images in the processed FP sample. When processing one or more FP images, after training, in addition to the training-based parameters of the descriptor generator 111 and the classifier 112, the PMC may also use predefined parameters and / or parameters received from other sources.

[0046] The FP images in the FP sample can come from different inspection modalities (e.g., from different inspection tools, from different channels of the same inspection tool (e.g., bright field and dark field images), from the same inspection tool using different operating parameters, or can be derived from design data, etc.).

[0047] For example, the FP image may be selected from the following: an image of a sample (e.g., a wafer or portion thereof) captured during a manufacturing process, a derivative of a captured image obtained through various pre-processing stages (e.g., an image of a portion of a wafer or photomask captured by a SEM or optical inspection system, an SEM image roughly centered around a defect to be classified by the ADC, an SEM image of a larger area where a defect to be localized by the ADC is located, registered images of different inspection modalities corresponding to the same mask position, segmented images, height map images, etc.), an image based on computer-generated design data, etc. It should be noted that the FP image may include an image of a layer of interest and / or registered images of one or more other layers of the sample. In the following, FP images of different layers are also referred to as images received from different modalities.

[0048] It should be noted that in embodiments of the presently disclosed subject matter, the characteristics of the images included in the FP samples and / or corresponding training samples are different from the conventional RGB images used in general deep neural networks known in the art. For example, electron-based imaging produces grayscale images with various effects such as uneven noise distribution, charging effects, large variability between sensors (different tools), etc. In addition, SEM images are usually composed of 5 different grayscale images, each corresponding to a different viewing angle (top, left, right, top, bottom) from which the image was taken.

[0049] It should also be noted that the FP samples and / or corresponding training samples and / or FP input data may include multiple data types (such as, for example, images of different sources and resolutions (e.g., defect images, category images, reference images, CAD images, etc.)), different types of numerical data (such as, for example, different types of data derived from images (e.g., height maps, defect masks, levels, segmentation, etc.)), different types of metadata (e.g., imaging conditions, pixel size, etc.), different types of manually constructed attributes (e.g., defect size, orientation, background segments, etc.), etc. A defect of a given category may come from one or more layers and / or one or more products in the FAB. The training set may be further enriched to include expanded defects and / or synthetic defects. As a non-limiting example, the training set may be enriched as detailed in PCT application No. PCT / IL2019 / 050150 filed on February 7, 2019 and incorporated herein by reference in its entirety, and as detailed in U.S. application No. 16 / 280,869 filed on February 20, 2019 and incorporated herein by reference in its entirety, etc.

[0050] refer to Figure 3, showing a general flow chart of a setup step (201). According to certain embodiments of the presently disclosed subject matter, the setup step includes: defining (301) a set of physical properties of defects (e.g., up to 20-30 physical properties) that can be used to distinguish between different classes of defects, and further defining (302) a set of values ​​of these physical properties for each given class, which uniquely describes the given class. Thus, each class is uniquely associated (303) with a multi-label data set corresponding to a corresponding unique set of values ​​of the physical properties that characterize the defects in the given class. Operations (301)-(303) can be provided for each product, for each customer, and / or globally for groups of products / customers, and the corresponding unique associations of the classes with the multi-label data sets can be stored in the PMC 104. Operations (301)-(303) can be provided manually or at least partially by a computer.

[0051] As non-limiting examples, the physical property may characterize physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, etc. Likewise, the physical property may characterize the location of a defect on certain layer or layers, chemical composition (e.g., absence and / or presence of a certain material), etc. The value of the physical property may be derived from images in the FP sample and / or corresponding training samples (e.g., by processing the image and / or its derivatives).

[0052] For example, defects in the "particle" category are characterized by a rough texture with clear edges (small z variation in the defect area) located on top of a pattern, while defects in the "bridge" category are characterized by a non-rough texture with no clear edges located on top connecting two patterns. Therefore, for these two categories, a set of physical properties of the defects can be defined as follows:

[0053] - Roughness of the texture

[0054] -Edge clarity

[0055] - Position relative to the top of the pattern

[0056] - Position relative to the two patterns

[0057] Optionally, but not necessarily, the values ​​of the physical properties may be defined in binary form. Table 1 presents a non-limiting example of a set of binary values ​​of physical properties that uniquely describe the "particle" and "bridge" categories of defects.

[0058] Table 1

[0059]

[0060] Thus, for the above example, the "particle" class may be uniquely associated with the multi-label binary dataset [1 1 1 0], and the "bridge" class may be uniquely associated with the multi-label binary dataset [0 0 1 1].

[0061] It should be noted that some attributes from a set of physical attributes may not be relevant to a particular defect class. Optionally, but not necessarily, values ​​for a physical attribute may be defined as "yes", "no" and "not relevant".

[0062] The PMC 104 trains the descriptor generator 111 to generate (304) a multi-label output vector (hereinafter also referred to as a "descriptor") for each given defect, wherein the multi-label output vector defines attribute values ​​from a set of physical attributes. As a non-limiting example, the training process can be provided in a "structured prediction" manner (e.g., as described in Nataly Brukhim and Amir Globerson, "Predict and Constrain: Modeling Cardinality in Deep Structured Prediction", published on February 13, 2018, arXiv:1802.04721v1; https: / / arxiv.org / pdf / 1802.04721.pdf. This article is incorporated herein by reference in its entirety)

[0063] Alternatively or additionally, the PMC 104 trains the classifier 112 using a data set uniquely associated with a corresponding class to provide ( 305 ) a classification of the defect based on a multi-label output vector (descriptor).

[0064] The training process produces a trained descriptor generator and a trained classifier.

[0065] Return to reference Figure 2 During runtime, the PMC 104 processes the FP samples using the trained descriptor generator to obtain descriptors that provide information of physical properties of the corresponding defects, and further uses the trained classifier to classify the defects based on a multi-label dataset uniquely associated with the corresponding class.

[0066] Classification may also include a deterministic process that defines a deterministic threshold as the ratio between the number of values ​​that match a given class and the total number of attribute values ​​in the multi-label dataset. Such a deterministic threshold enables optimization of the confidence level of defect classification and identification of misclassified defects. As a non-limiting example, such a threshold may be implemented in a technique for classifying defects, comprising assigning each class to a classification group in three or more classification groups with different priorities, and further separately setting purity, accuracy, and / or extraction requirements for each class, and optimizing the classification results according to the requirements for each class. The technique is disclosed in U.S. Application No. 2019 / 0096053, which is incorporated herein by reference in its entirety.

[0067] Analysis of the generated descriptors can identify new recurring attribute patterns, thereby enabling detection of new categories that were not initially included in the classification. Likewise, such analysis enables identification of multimodal behavior of one or more categories and definition of corresponding subcategories (e.g., corresponding to different clusters of values ​​of attributes characterizing the same category).

[0068] It should be understood that the application of the present invention is not limited to the details set forth in the description contained herein or shown in the accompanying drawings. The present invention can have other embodiments and can be practiced and implemented in various ways. Therefore, it should be understood that the wording and terminology adopted herein are for illustrative purposes and should not be considered as limiting. Therefore, those skilled in the art will recognize that the concepts on which the present disclosure is based can be easily used as the basis for designing other structures, methods and systems for achieving several purposes of the subject matter currently disclosed.

[0069] It should also be understood that the system according to the present invention can be implemented at least in part on a suitably programmed computer. Likewise, the present invention contemplates a computer program readable by a computer for performing the method of the present invention. The present invention further contemplates a non-transitory computer readable memory tangibly embodying a program of instructions executable by a computer for performing the method of the present invention.

[0070] It will be readily appreciated by those skilled in the art that various modifications and changes may be applied to the embodiments of the invention as hereinbefore described without departing from the scope of the invention as defined in and by the appended claims.

Claims

1. A method for automatically classifying defects in a semiconductor sample into a plurality of categories, the method comprising, by a processing and memory circuit system PMC: obtaining a set of physical properties of the defect, the set of physical properties being usable to distinguish different classes of defects from among the plurality of classes; obtaining data providing information of multi-label data sets, each data set uniquely indicating a respective category of the plurality of categories and including a unique set of values ​​for a physical property from the set of physical properties; as well as A first machine learning model is trained using the data to process a sample comprising one or more images providing information of a given defect so as to generate, for the given defect, a multi-label output vector providing information of a value of the physical property from the set of physical properties, thereby generating a multi-label descriptor for the given defect.

2. The method of claim 1, further comprising: A second machine learning model is trained using the data to provide multi-label classification.

3. The method of claim 1, further comprising: The generated multi-label descriptors of the defects in the samples are analyzed to identify new repeated multi-label datasets to identify new categories of the defects.

4. The method of claim 1, further comprising: The generated multi-label descriptors of the defects in the samples are analyzed to identify one or more categories of multimodal behavior.

5. The method of claim 2, wherein providing a multi-label classification of the given defect comprises: A certainty threshold is defined as the ratio between the number of values ​​in the multi-label descriptor that match a given class and the total number of values ​​in the multi-label dataset that indicate the given class.

6. The method of claim 5, further comprising: Using the certainty threshold enables at least one of the following: a. Optimize the confidence level of defect classification; b. Identify the defects of misclassification; c. Set purity requirements for each category and / or group of categories; d. setting accuracy requirements for each category and / or group of categories; and e. Set extraction requirements for each category and / or group of categories separately.

7. The method of claim 1, wherein the physical attribute in a set of physical attributes of the defect provides information of at least one of: physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, layer, texture, edge, and chemical composition.

8. The method of claim 1, wherein the plurality of categories include a "particle" category and a "bridge" category, and wherein the set of physical properties of the defect includes roughness of texture, sharpness of edge, position relative to the top of a pattern, and position relative to two patterns.

9. The method of claim 2, wherein the multi-label descriptor and the values ​​of the physical attributes in the multi-label dataset are binary.

10. The method of claim 2, wherein the multi-label descriptor and the values ​​of the physical attributes in the multi-label dataset correspond to "yes", "no" and "not relevant" with respect to relevant physical attributes.

11. A system for classifying defects in a semiconductor sample into a plurality of categories, the system comprising a processing and memory circuit system (PMC) operably connected to an input interface, wherein the input interface is configured to receive a sample including an image providing information of the defects; and wherein the PMC is configured to: using the trained first machine learning model to generate a multi-label descriptor of the defect in the semiconductor sample, the multi-label descriptor being usable for classification, wherein the first machine learning model is trained using data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category from the plurality of categories and comprising a unique set of values ​​of a physical property from a set of physical properties of the defect, the first machine learning model being trained to process samples comprising one or more images providing information of a given defect so as to generate a multi-label output vector providing information of the values ​​of the physical property from the set of physical properties for the given defect, thereby generating a multi-label descriptor for the given defect.

12. The system of claim 11, wherein the PMC is further configured to: After training a second machine learning model to provide multi-label classification, the trained second machine learning model is used to classify defects in the semiconductor sample by matching the respectively generated multi-label descriptors of the defects with the multi-label dataset.

13. The system of claim 11, wherein the PMC is further configured to: The generated multi-label descriptors of the defects in the semiconductor sample are analyzed to identify new repeating multi-label data sets to identify new classes of the defects.

14. The system of claim 11, wherein the PMC is further configured to identify one or more categories of multimodal behaviors by analyzing the generated multi-label descriptors of the defects in the samples.

15. The system of claim 12, wherein classifying the defects of the defects in the semiconductor sample comprises: A certainty threshold is defined as the ratio between the number of values ​​in the multi-label descriptor that match a given class and the total number of values ​​in the multi-label dataset that indicate the given class.

16. The system of claim 15, wherein the PMC is further configured to use the certainty threshold to enable at least one of: a. Optimize the confidence level of defect classification; b. Identify the defects of misclassification; c. Set purity requirements for each category and / or group of categories; d. setting accuracy requirements for each category and / or group of categories; and e. Set extraction requirements for each category and / or group of categories separately.

17. The system of claim 11, wherein the physical attribute in the set of physical attributes of the defect provides information of at least one of: physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, layer, texture, edge, and chemical composition.

18. The system of claim 11, wherein the plurality of categories include a "particle" category and a "bridge" category, and wherein a set of physical properties of the defect include roughness of texture, sharpness of edge, position relative to the top of a pattern, and position relative to two patterns.

19. A non-transitory computer readable medium comprising instructions which, when executed by a computer, cause the computer to perform a method of classifying defects in a semiconductor sample into a plurality of categories, the method comprising: using the trained first machine learning model to generate a multi-label descriptor of the defect in the semiconductor sample, the multi-label descriptor being usable for classification, wherein the first machine learning model is trained using data providing information of a multi-label dataset, each dataset uniquely indicating a corresponding category from the plurality of categories and comprising a unique set of values ​​of a physical property from a set of physical properties of the defect, the first machine learning model being trained to process samples comprising one or more images providing information of a given defect so as to generate a multi-label output vector providing information of the values ​​of the physical property from the set of physical properties for the given defect, thereby generating a multi-label descriptor for the given defect.

20. The non-transitory computer readable medium of claim 19, wherein the method further comprises: After training a second machine learning model to provide multi-label classification, the trained second machine learning model is used to classify defects in the semiconductor sample by matching the respectively generated multi-label descriptors of the defects with the multi-label dataset.

Images