WO2003100405A1

WO2003100405A1 - Defect classification device generation method and automatic defect classification method

Info

Publication number: WO2003100405A1
Application number: PCT/JP2003/006353
Authority: WO
Inventors: Atsushi Miyamoto; Hirohito Okuda; Toshifumi Honda; Yuji Takagi; Takashi Hiroi
Original assignee: Hitachi High-Technologies Corporation
Priority date: 2002-05-23
Filing date: 2003-05-21
Publication date: 2003-12-04

Abstract

In order to generate a defect classification device satisfying a classification request from a user, firstly, the defect classification device is expressed by a decision tree for hierarchically classifying defect classes by a plurality of branching elements. Separate classification rules are assigned to the respective branching elements. Configuration of the decision tree and specification of the classification rules are determined according to display of distribution state of attributes of defective samples (including separation degree of each attribute between defective samples belonging to each defect class) indicated by the user for each of the branching elements. Moreover, the classification request from the user is clarified by using GUI which almost simultaneously displays various inspection information obtained from a plurality of defect inspection devices.

Description

Description Defect classifier generation method and defect automatic classification method

The present invention relates to a method for generating a defect classifier for automatically classifying foreign objects and defects generated on a sample such as a semiconductor wafer in a semiconductor manufacturing process, and to automatically classify defects using the generated defect classifier. It relates to an automatic defect classification method and its system. Background technology,

A semiconductor device is manufactured by performing a plurality of processes such as exposure, development, and etching on a substrate to be a substrate, and after being processed in a predetermined processing step among the plurality of processing steps, an optical or semiconductor device is manufactured. Defect position and size are inspected using SEM (Scanning Electron Microscope) type foreign matter inspection equipment ゃ panning inspection equipment. Although the number of detected defects depends on the state of the manufacturing process, it can range from hundreds to thousands per wafer, and high-speed defect detection is required for these defect inspection systems. Hereinafter, the defect inspection devices that perform these defect detections are collectively referred to as defect detection devices.

On the other hand, after inspection by the defect detection device, it is possible to perform a detailed re-inspection of the defect detected by the defect detection device using an optical or SEM type defect review device with a higher imaging magnification. is there. However, it is not practical to perform a detailed inspection of all the defect samples detected by the defect detection device due to time constraints, and a set of defects detected by the defect detection device is sampled, and a detailed inspection of the subset is performed. Done. Hereinafter, these defect inspection devices that perform defect review inspections are collectively referred to as defect review devices. Call. Defect detection device and defect review device, scanning probe microscope

(Scanning Probe Microscope: SPM) and inspection equipment such as elemental analysis equipment are collectively referred to as defect inspection equipment.

In the defect review device, a function to automatically obtain an enlarged image of the defect based on the defect position information from the defect detection device, that is, the ADR (Automatic Defect Review) function, Products with a function to automatically classify defect types based on detailed information such as pod shape and texture (surface pattern), that is, an ADC (Automatic Defect Classification) function are being developed. On the other hand, some defect detection devices have a simple defect coarse classification that emphasizes high-speed processing, that is, a classification function called RTADC (Real Time-ADC).

As for the method of automatically classifying defects based on various kinds of inspection information, various methods have been studied since ancient times as multivariate analysis methods in the field of computer recognition.

One of the classical methodologies is a method called rule-based classification. In this methodology, various image features are extracted from the images to be classified, and the values of the image features are determined based on the rules of the “if-then” formula built into the system, thereby detecting defects. Classify into one of the classes. In the rule-based classification, the defect classes and rules to be classified are fixed and cannot respond flexibly to the needs of the user, but they do not require teaching data, so they can be used from the start of the production process. There is an advantage.

Another classical methodology is a method called learning-type classification. In this methodology, teacher images are collected in advance and learned to optimize the classification rules (eg, neural nets). Learning-based classification may be able to perform flexible classification according to the user's requirements, but generally it is necessary to collect a large amount of teaching data in order to obtain good performance. There is a problem that it cannot be used practically when starting up the production process. Conversely, when only a small number of teaching data is used, it is known that the learning is over-adapted to the teaching data, which is called over-learning, and the performance is reduced.

Japanese Patent Application Laid-Open No. 2001-135692 discloses an automatic defect classification method that can be applied uniformly in a hybrid as a configuration using both the rule-based classification and the learning classification described above. .

Further, as the prior art relating to defect classification, JP-A-11-344450, JP-A-2001-93950, JP-A-201-127271, JP-A-2,001-2. JP-A-56480, JP-A-2001-331817, JP-A-2002-14054 and JP-A-2002-90312 are known. Disclosure of the invention

However, even if the classification result output by the defect-based classifier composed of the rule-based classification, the learning-based classification, or a combination thereof does not match the classification request of the user, the classification inside the system is performed. It is not easy to correct standards. In other words, in the rule-based classification, if the meaning of various attributes used as a classification criterion is unclear, the user performs customization such as selection of various attributes and threshold setting for his / her own classification request. It is difficult. Also, in the case of automatically generating classification rules in learning-type classification, if many attributes are carelessly prepared as feature values, the degree of freedom of learning increases, and overfitting with a small number of teaching There is a risk of over-learning. Therefore, more teaching samples are required. A similar problem exists in a configuration in which rule-based classification and learning-type classification are combined. The issue is how to determine the appropriate configuration.

The attributes of the defect obtained from the defect inspection device include information on the image feature amount, defect coordinates, composition analysis results, construction history, information on the distribution of defect positions detected on the equipment QC (Quality Control) or wafer, and When the number of defects is listed, and attributes obtained from a plurality of heterogeneous defect inspection devices such as an optical or SEM foreign particle inspection device, pattern inspection device, defect review device, SPM, and elemental analysis device can be referenced. There is also. Although the automatic defect classification is performed using the above-described attributes as a criterion, it is a major issue for the user how to use these large number of attributes and generate a defect classifier that satisfies an expected defect classification standard.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-described problems and to provide a defect that clarifies a user's classification requirement, generates a defect classification class, and classifies the defect into the defect classification class based on appropriate user support. It is another object of the present invention to provide a method for generating a classifier and an automatic defect classification method using the generated defect classifier, and a system therefor. To provide a defect classifier including a defect classification class for each defect inspection device, a defect automatic classification method and its system, and a method for obtaining data consistency between inspection devices ゃ a data interpolation method Is to do.

In order to generate a defect classifier that satisfies the user's defect classification requirement according to the present invention, first, it must be determined whether or not various attributes are valid as criteria for the defect classifier. As a mechanism for visually associating a user's classification request with the degree of effectiveness of various attributes with respect to the classification request, the above-mentioned attributes may be independently or combined as necessary, or may be converted arbitrarily (arbitrary plural attributes). A GUI (Graphic User Interface) with the function of visualizing the distribution after adding principal component analysis, compression processing of the number of dimensions of attributes, and arbitrary conversion processing using force channel functions, etc. is provided. Also, The attribute is at least one of image feature quantity, defect classification result, defect coordinate, composition analysis result, arrival history, apparatus QC, and information on the distribution of defect positions detected on the wafer and the number of defects. It includes the above.

On the other hand, the defect classifier according to the present invention is configured by a decision tree that classifies defect classes hierarchically by a plurality of branches. The user can grasp the degree of separation of each attribute between the defect samples belonging to each defect class divided at an arbitrary branch in the decision tree by the distribution state of the visualized various attributes. The degree of separation in various attributes is one of the indicators for judging whether the information is effective as a classification criterion in each branch, and based on such information, various attributes are selectively or comprehensively utilized. It decides a decision tree, selects a classification rule, or controls parameters. The contents of these processes are determined based on the defect classification request of the user.

Further, in order to clarify the user's classification request, the present invention provides a GUI which acquires and shares inspection information including defect images and attributes in various defect inspection apparatuses and has a simultaneous review screen of the inspection information. . The user can comprehend the inspection information comprehensively based on the GUI and clarify his / her classification request.

Further, in the present invention, in combination inspection using a plurality of defect inspection devices, for example, a defect detection device and a defect review device, first, inspection information obtained from both inspection devices is selectively or collectively used. Generate a defect classifier that includes the defect classification class that meets the user's classification requirements. Next, using only the inspection information obtained from the defect detection device, a defect classifier including a defect classification class in which the defect classification class meeting the user's classification request is a subset or a similar set is set. The generated defect classifier including the defect classification class is employed in the defect detection device. In this way, by setting the defect classification class of the defect detection device, it is possible to obtain hierarchical consistency between the defect classification classes. This enables effective defect classifier learning and effective review sampling in a defect review device that is performed later.

Further, in the present invention, the same or similar type of defect sample may have different determination criteria among a plurality of defect inspection apparatuses in an attribute obtained from an arbitrary processed image. As a matching method in such a case, a teacher pattern is generated from a processed image or an artificially generated image obtained from an arbitrary defect inspection device or a CAD data obtained from a user who has obtained a result expected by a user. By adjusting the image processing procedure or image processing parameters so that the processed images of the same defect image in the defect inspection apparatus match or become similar, the same attribute can be analyzed among a plurality of defect inspection apparatuses. It can be used as a criterion for the determination.

Further, according to the present invention, when a plurality of defect inspection apparatuses cannot refer to inspection information of a defect sample of interest in an arbitrary defect inspection apparatus due to a difference in inspected samples, a defect sample group in the defect inspection apparatus is referred to. Classify into several classes based on the spatial distribution of defects, defect classification results, or various inspection information. Spatial membership belonging to the above class is defined for arbitrary coordinates on the defect distribution map, and common or similar inspection information for the same class is used commonly for defect samples in the same class. Alternatively, the inspection information of the defect sample of interest can be interpolated by selectively using the inspection information of the defect samples in the same class. Further, in order to achieve the above object, the present invention provides an inspection information acquiring step of inspecting a defect sample group on an arbitrary sample by at least an arbitrary defect inspection device to acquire sample inspection information; A display step of displaying the state of the defect attribute distribution of the defect sample group on the arbitrary sample based on the sample inspection information obtained in the inspection information obtaining step, and a display step of displaying the displayed defect attribute distribution on the screen. Classification of defect samples based on condition A decision tree setting step including a classification rule setting step of setting an individual classification rule for each branch element in a decision tree in which class elements are hierarchically expanded through branch elements. The present invention also provides a multi-dimensional method in which a plurality of attributes determined by a group of teaching defect samples belonging to a category classified for each branch element in the decision tree are combined as the state of the defect attribute distribution. It is characterized by a graph.

Further, the present invention provides an inspection information acquiring step of inspecting a defect sample group on an arbitrary sample with a plurality of defect inspection devices to acquire a plurality of sample inspection information, and a plurality of samples acquired by the inspection information acquiring step. Inspection information is displayed almost simultaneously on the screen, and this is browsed to determine the classification class element of the defect sample group. On the screen, the classification class element is branched based on the classification class element determined by the reuse step. A decision tree designating step of designating a decision tree hierarchically expanded via elements, and the designated decision based on at least one sample examination information acquired by the examination information acquiring step on a screen A classification rule setting step of setting an individual classification rule for each branch element in the tree. A method of generating a defect classifier and having a flop.

Further, the present invention classifies the defects based on inspection information obtained by inspecting a defect generated in an inspection target in a second defect inspection device in which inspection is performed after inspection by the first defect inspection device. A method of generating a second defect classifier for, wherein the second defect classifier is obtained by inspecting a defect sample group on an arbitrary sample by the first and second defect inspection devices It is characterized in that it is generated or changed using part or all of the sample inspection information. In addition, the present invention provides a first defect inspection apparatus in which an inspection is performed before an inspection by a second defect inspection apparatus, the inspection being performed by inspecting a defect generated in an inspection target. A first defect classifier for classifying the defect based on inspection information, wherein the defect classification class classified by the second defect classifier in the second defect inspection apparatus is the first defect classifier. The first defect classifier is generated or changed so as to be a subset of defect classification classes classified by one defect classifier or a classification similar thereto.

Also, the present invention provides an inspection information acquisition step of inspecting a defect sample group on a sample by at least a combination of the first and second defect inspection devices to acquire first and second inspection information; Generating a first or second defect classifier according to the inspection order of the first and second defect inspection devices with reference to the first and second inspection information acquired in the step; A defect classifier generation method comprising the steps of: BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a block diagram showing an embodiment of the configuration of a data server for sharing inspection information and various defect inspection devices according to the present invention.

FIG. 2 is a flowchart showing one embodiment of a process of acquiring various kinds of inspection information in various kinds of defect inspection devices according to the present invention.

FIG. 3 is a flowchart showing one embodiment of a method for generating a defect classifier according to the present invention.

FIG. 4 is a view showing an embodiment of a simultaneous review of inspection information, a distribution display of various attributes of defects, and a list display window of defect classes and branches according to the present invention.

FIG. 5 is a diagram showing an embodiment of a decision tree structure setting and defect sample teaching window according to the present invention.

FIG. 6 is a diagram showing an embodiment of a procedure for setting a decision tree structure in FIG. FIG. 7 is a diagram showing an embodiment of a classification rule setting window for each branch in the decision tree according to the present invention.

FIG. 8 is a diagram showing an embodiment of a multidimensional graph display of various attributes according to the present invention and a method of specifying a constraint condition in the graph.

FIG. 9 is a defect distribution map showing an embodiment of the defect sample distribution and the automatic defect classification result in each step of the defect classifier generation procedure according to the present invention.

FIG. 10 is a flowchart showing one embodiment of a defect classifier setting procedure and a defect classification procedure according to the present invention.

FIG. 11 is a diagram showing the relationship between defect classification classes in each defect inspection apparatus as shown in the defect distribution map in FIG. 9 according to the present invention.

FIG. 12 is a diagram showing an example of a class distribution generated when interpolating missing test information according to the present invention.

FIG. 13 is a diagram showing an embodiment of a procedure for matching inconsistencies between inspection information of a plurality of defect inspection apparatuses according to the present invention. BEST MODE FOR CARRYING OUT THE INVENTION

An embodiment according to the present invention will be described with reference to the drawings. The present invention presupposes that at least one or more defect inspection apparatuses can acquire inspection information and that the inspection information can be referred to from any defect inspection apparatus. The defect inspection apparatus 101 may be an optical or SEM type foreign object inspection apparatus, a pattern inspection apparatus, a review inspection apparatus, an element analysis apparatus, or another type of defect inspection apparatus. Includes inspection equipment with different inspection processes, or defect inspection equipment of the same or different type, regardless of the process. FIG. 1 is a block diagram showing an embodiment of an automatic sample defect classification system according to the present invention and a configuration of each defect inspection apparatus used in the system, wherein 101a to: 101n. The defect inspection devices A to N are the processing terminal devices that process the inspection information from each of the defect inspection devices 101a to; Reference numeral 107 denotes a data server, and reference numeral 108 denotes a processing terminal device that processes information from the data server 107. Information can be transmitted and received between 101, 102, 107 and 108 via the network 103. Inspection information from each of the defect inspection devices 101a to 101n 'is processed by the processing terminal devices 102a to 102n' and managed by the data server 107, shared or shared with other processing terminals 102a to 102n. '.' Or directly from the processing terminal device 108 of the entire system (for the data server). The inspection information is viewed or processed and analyzed at the same time as or after the inspection in the processing terminal devices 102 a to 102 n for any defect inspection device or the system processing terminal device 108. .

Next, a method of acquiring inspection information in the arbitrary defect inspection apparatus 101 and the processing terminal apparatus 102 and a type thereof will be described in detail. After processing in a predetermined processing step of semiconductor device manufacturing, a defect inspection is performed by an arbitrary defect inspection apparatus 101, and various inspection information (various defect classification results) is obtained from the processing terminal apparatus 102 or. FIG. 2 shows the details of this processing. That is, FIG. 2 is a flowchart showing an embodiment of the processing operation in the processing terminal device 102 or 108 for acquiring various inspection information (various defect classification results) using an arbitrary defect inspection device. In step 201, first, an arbitrary defect inspection apparatus 101 detects an arbitrary defect position on an inspection target such as an actual wafer and a sample inspection target (a sample for generating the defect classifier 120). Then, a stage (not shown) is moved to a defect position and a reference position, and an image signal is detected and an image signal is detected. The image signal is converted into a digital image signal, and the converted digital image signal is provided to the processing terminal device 102 or 108. In step 203, the processing terminal device 102 or 108 calculates various image feature amounts of the defect by comparing, for example, the captured defect image with the reference image. The processing of each step in the loop 204 encircled by the dotted line is repeated in the same manner at any other defect position, and the processing terminal device 102 or 108 integrates them in step 205. To obtain defect distribution information on the sample. However, in the defect review device, the defect detection in step 201 is performed by inputting defect position information detected by an external foreign matter inspection device or pattern inspection device in advance. Next, in step 206, inspection equipment information (such as optical, SEM, AFM (atomic force microscope), etc.) that is unique to each defect inspection equipment, such as equipment QC (Quality Control) and construction history Information, information on resolution and sensitivity, etc. or wafer information specific to the wafer (information on the manufacturing process to be inspected and information on the circuit structure (memory area, logic area, etc.)). In step 207, the processing terminal device 102 or 108 determines the defect distribution information on the sample obtained in step 205, the image feature amount for each defect obtained in step 203, and step 2. Based on inspection information obtained by combining inspection device information and wafer information, a generated defect classifier 120 according to the present invention obtained, for example, by the processing flow shown in FIG. 3, for example, by a plurality of branches Defect classification is performed using a decision tree that classifies defect classes hierarchically. Here, the inspection information includes at least one of various defect images and various attributes, and the various defect images refer to other defect inspection devices regardless of whether a defect is actually detected. In addition to all the defect images taken by the detectors of each defect inspection apparatus 101 including the defect image and the reference image, arbitrary image processing (for example, binarized image processing, dilation, Shrinkage image processing) It contains at least one or more of the processed images. The defect image and the reference image are images obtained by moving the stage to the defect position detected on the wafer and the reference position, respectively, and the image is taken. On a chip different from an existing chip (for example, an adjacent chip), it refers to a position corresponding to the defect position. However, there is a technique for synthesizing a pseudo-reference image from a defect image using the partial periodicity of the defect image, and is generally referred to as a reference image (a reference image for comparison with the defect image). Further, the processed image includes a processed image obtained by synthesizing at least two or more arbitrary image groups by arbitrary image processing. The various attributes include an image feature amount, a defect classification result (result of which defect class is classified into an arbitrary defect sample), a defect coordinate, and a composition analysis result (combination of any composition of the arbitrary defect sample). At least one of the following: information on the start of construction, history of construction, equipment QC, information on the distribution of defect positions detected on the wafer, and the number of defects (hereinafter referred to as defect distribution information). It is a thing. In addition, the image feature amount refers to the color (texture, etc.), size (indicated by area or length), shape (foreign matter shape, flaw shape, etc.), or wiring pattern of a defect obtained from the various defect images. Quantitative characteristics such as the positional relationship of defects (positional relationship indicated by fatal defects such as short-circuit defects and disconnection defects), or any heterogeneous defect inspection device from defect images obtained by any defect inspection device This includes all or some of the newly designed and calculated image features that are considered effective in.

In the above description, the defect inspection apparatus 101 executes up to step 202 including the actual inspection target and the sample inspection target (sample for generating the defect classifier 120). Although the subsequent processing and the generation of the defect classifier 120 have been described as being performed by the processing terminal devices 102 and 108, the defect inspection device 101 performs steps 204, 205 and 2 06 Inspection information (actual inspection target and And sample inspection target) are obtained, and a defect classifier is used based on the inspection information.

The processing of generating the 120 and performing the actual defect classification based on the generated defect classifier 120 may be executed by the processing terminal devices 102 and 108. 'It is necessary that the defect classifier 120 according to the present invention be generated before classifying a defect generated on an actual inspection target.

[First Embodiment] (Generation of a defect classifier 120 according to the present invention in a defect inspection apparatus)

The first embodiment will be described. In the present embodiment, at least one or more defect inspection apparatuses 101 are used for automatic classification of foreign matter and defects on a semiconductor wafer (hereinafter collectively referred to as “defects” unless otherwise specified). The method for simultaneously reviewing inspection information and the method for generating the defect classifier 102 according to the present invention employed in any one of the defect inspection apparatuses 101 or the processing terminal apparatuses 102 and 108 are greatly increased. Consists of two. In the following description, the generation of the defect classifier 102 is performed by one of the processing terminal devices 102 a to 102 η or the processing terminal device 108. For this purpose, in FIG. 1, each of the processing terminal devices 102 and 108 has, in each of the processing terminals 102 and 108, a calculating means having a function of a GUI 11 ° and a function of generating a defect classifier 120, It is provided with a storage means 131, a display means 132, an input means 133, and the like connected to the computer.

1.1 Processing flow

The outline of the process in the present invention will be described with reference to FIG. First, acquisition of various kinds of inspection information for generating the defect classifier 120 from the various defect inspection apparatuses 101 and the processing terminal apparatus 102 (step 301) has already been performed. You can refer to this. However, the information that can be referred to here is not limited to the inspection information obtained from one defect inspection device 101 as described above, but may be an optical or SEM type defect inspection device, a pattern defect inspection device, a SEM. Formal When a defect inspection apparatus 101 such as any defect review apparatus or elemental analysis apparatus has performed a defect inspection, the defect classifiers 120 obtained from the plurality of defect inspection apparatuses 101 are used. Inspection information for generating the defect inspection information can also be referenced and used as needed, in which case, as shown in FIG. 1, various defect inspection devices, for example, defect inspection devices A (101a), B ( The inspection information from 101b) is processed by the processing terminal devices A (102a) and B (102b) and managed by the data server 107 using the network 103 etc. And shared system configuration. That is, with the system configuration, the processing terminal device 102 or 108 can acquire and share inspection information including defect images and attributes in various defect inspection devices in step 301. . Next, in the processing terminal device 102 or 108, based on the inspection information, a user clarifies a defect classification standard considered to be an ideal defect classification based on system support (step 302). Then, a defect classification class is created (step 303). The details will be described later. After the defect classification class is created in the processing terminal device 102 or 108, the process proceeds to the step of determining a defect classifier in the system that realizes the defect classification. The defect classifier is represented by a decision tree that classifies defects hierarchically by a plurality of branches, and the design of the defect classifier 120 according to the present invention is completed by setting individual classification rules in each branch described above. I do.

The procedure is composed of various steps shown in a loop 304 surrounded by a dotted line. The loop 304 basically consists of five steps: “Decision tree decision for defect classification (Step 300)” “Teach defect sample (Step 300)” “Evaluation of the degree of classification of defect attribute distribution (Step 300) ) ”,“ Selection of classification rules (Step 308) ”, and“ Evaluation of classification results (Step 309) ”. However, of these steps, steps 303, 307, and 309 can be freely skipped if not required. For example, to determine the classification rules For this reason, if it is not necessary to refer to the distribution of the various attributes of the defect, steps 306 and 307 are unnecessary. It is also conceivable to proceed to the next decision tree without performing step 309 for evaluating the classification results. In addition, the above five steps can be performed by changing the order as necessary. For example, it is possible to first teach defect samples for all defect classes, or to determine a decision tree based on the degree of separation of various attributes of a defect. Furthermore, some or all of the above five steps can be automated or semi-automated. For example, the function has a function of quantifying the degree of separation of the attribute distribution in step 307, and automatically determining an appropriate decision tree or classification rule based on the degree of separation, and selectively adopting the function. can do. Alternatively, for example, regarding the decision tree for the defect classification (step 300), some patterns have already been registered in the system data base, and the data is selected from the data set in the data base. The decision can be made with reference to the overnight base.

In the processing terminal device 102 or 108, the loop 304 is performed a plurality of times as necessary, and when the decision tree is determined (step 310), the defect classifier (classification rule) 120 Generation ends. Next, each step will be described in detail. First, the defect classification class generation (user determination of defect classification standard) (step 303) based on the system support by simultaneous review (step 302) will be specifically described.

1.2 Determining defect classification class by simultaneous review (Steps 302, 303)

First, there is a step 303 for visually performing defect classification on some defect samples and determining a defect classification class. This is a step in which the user decides what kind of defect group and what kind of defect group he or she wants to be classified as a homogeneous or heterogeneous defect class. Determine defect classification class To do so, the user must first clarify his own defect classification criteria. Also, due to the classification performance, it is not always possible to design a defect classifier that completely satisfies the user's classification criteria. In addition, inconsistencies may occur between multiple pieces of inspection information (for example, one inspection information may be judged to be a foreign matter defect and the other inspection information may be judged to be a false alarm. Inconsistent), the user may be required to make a unified decision regarding defect classification. In consideration of the above, it is effective to browse the inspection information from multiple defect inspection devices at the same time to determine the defect classification class (step 302), and to check the inspection information for that purpose. Provide a simultaneous review screen to the user. Window 400 shows an example of a GUI (Graphic User Interface) 110 having a simultaneous review screen. As described above, by providing the GUI having the simultaneous review screen of the acquired and shared inspection information, the user can comprehensively grasp the inspection information based on the GUI 110, and Classification requirements can be clarified.

There are many cases where the integrated utilization of the multiple inspection information is effective for defect analysis and classification. For example, SEM-based defect inspection equipment has difficulty in observing defects in the lower layer of the wafer, while optical-type defect inspection equipment can observe lower-level defects relatively well. On the other hand, with respect to VC (Voltage Contrast) defects, which are electrical defects existing in the lower layer of the wafer, SEM type defect inspection equipment can observe better than optical defect inspection equipment. Further, it is known that a defect detector having a large field of view (coarse resolution) tends to be able to observe a defect better with respect to the VC defect and the like. As described above, the inspection information obtained from each of the defect inspection apparatuses has advantageous and disadvantageous aspects depending on the type of the defect. Even for inspection information obtained from equipment, a wide variety of information obtained by different detection methods and processing methods is used for defect analysis and classification. It is effective for

1.2.1 Simultaneous Review (Display of Defect Distribution Map) (Step 302) Hereinafter, a method of displaying a defect map will be described with reference to FIG. FIG. 4 is a diagram showing an embodiment of a simultaneous review of inspection information, a distribution display of various attributes of defects, and a list display window of defect classes and branches.

In order to obtain a unified view from multiple test information, it is effective to browse those test information at the same time. Therefore, in step 302 shown in FIG. 3, a screen for simultaneous review of inspection information is provided to the user. The window 400 in FIG. 4 shows an example of the GUI in the present system (for example, the processing terminal devices 102, 108) shown in FIG. Naturally, in order to realize the GUI 120, the present system (for example, the processing terminal devices 102, 108) interacts with the display device 132 through graphics and images on the display device. It comprises a computer (including storage means 13 1 for storing various data) (not shown) and input means 13 3 such as a keyboard. Furthermore, the computer also has a defect classifier 120 customized based on the inspection information according to the present invention.

Next, details of the information referred to in step 302 will be described. The GUI shown in the window 400 has a function of displaying a defect distribution map in an arbitrary defect inspection apparatus. If the inspection is performed in a plurality of defect inspection apparatuses, the GUI is displayed in the window 400. Has a function of displaying a defect distribution map in each defect inspection device side by side. Hereinafter, an example in which the inspection information of the two defect inspection apparatuses (defect detection apparatus A and defect review apparatus B) can be referred to will be described, but the present invention is not limited to this combination. Defect inspection equipment of different types, or defect inspection equipment of the same type that performs inspections in different steps, or any one of defect inspection equipment of the same type but different inspection equipment More than one Includes a combination of defect inspection devices. Since then, defect review equipment

We will deal with the problem of setting the defect classifier 120b in B, but we also consider the setting of the defect classifier 120 in any defect inspection device 101 or processing terminal device 102, 108. Can be

In this embodiment, reference numerals 408 and 409 denote defect distribution maps indicating defect distributions in the defect detection device A and the defect review device B, respectively. Reference numeral 420 denotes a defect when the defect classifier according to the present invention, which is customized based on the inspection information obtained from the defect detection device A and the defect review device B, is applied and reclassification is performed. 5 is a defect distribution map showing a distribution. By using the defect distribution map 420, it is possible to interactively customize the defect classifier while checking the classification result by the set defect classifier. By pressing the “map display method” button 4 17 and performing predetermined settings, the defect distribution results such as the defect classification result and various attributes of each defect sample can be displayed on the defect distribution map 4 0 8 and 4 0. A function to display two-dimensionally or three-dimensionally by displaying characters, numerical values, color coding, emphasis, etc. on 9, 420 is provided, and the user can grasp the whole image of the defect distribution. In addition, it has a function to display the device QC and the history of the start of construction for each defect inspection device. The check boxes 411 and 412 are for displaying the equipment QC, and the check boxes 411 and 413 are for displaying the history of the start of construction.

1.2.2 Display of defect image (Step 302)

The inspection information simultaneous review window 401 shown in FIG. 4 has a function of designating an arbitrary plurality of defect samples and displaying the defect images of the defect samples side by side so that they can be viewed almost simultaneously. As an example, consider two defect samples da 1 and da 2 in the defect distribution map 408 in the defect detection device A. These defects correspond to the defect samples db1 and db2 in the defect distribution map 409 in the defect review apparatus B, respectively. However, The defect samples da1 (dbl) and da2 (db2) are classified into the same class Ca2 in the defect detection device A, and are classified into different classes Cb3 and Cb2 in the defect review device B. ing. Here, the defect sample da 1 is selected, and is dragged, dropped, and dropped on the window 402 to display the defect image 404 of the defect sample da 1 acquired by the defect detection device A as shown in the figure. be able to. When the defect sample da 1 is displayed and the corresponding defect sample db 1 of the defect review apparatus B exists in the defect distribution map 409, the defect image 406 of the defect sample db 1 is automatically displayed in the window 403. This function displays the defect sample da 1 or db 1 regardless of whether you drag, drop, and drop it to windows 402, 403, respectively. Things. Similarly, by selecting either of the defect samples da 2 and db 2, the defect image 40 5.407 can be displayed in the windows 402 and 403. Defect images 404 From the simultaneous defect review images of 4 to 407, the user can determine whether the two defect samples da 1 (db 1) and da 2 (db 2) should be classified as different defect classification classes . At this time, as a mechanism to support the user, any of the defect images can be displayed by pressing the “image display method” button 416 and performing a predetermined setting. In addition, by pressing the "Detailed information" button 418, the inspection information list for any defect sample specified on the defect distribution map can be displayed simultaneously. <Also, by pressing the "Search" button 419 A similar defect search can be performed. The search is performed by specifying a search range such as a defect inspection apparatus or a region on a wafer and specifying a search formula expressed by arbitrary inspection information or a combination thereof.

The determination of the defect classification class in step 303 is based on the simultaneous This can be performed by visual observation of the user based on the user screen 401. However, if any rule is applied, for example, "When the defect is detected by the former and not detected by the latter when the inspection is performed using the SEM and optical defect inspection devices, the defect classification result is It is possible to automate some or all of the defect classification class decisions by introducing knowledge such as “VC defect is likely to be false information”. The same applies to the generation of the defect classifier according to the present invention.

■ 1.2.3 Defect classification class creation (Step 303)

Next, the designation of the defect class name and number performed in step 303 based on the user's ideal defect classification standard confirmed on the simultaneous review screen 401 will be described.

As an example, if the defect classification standard requested by the user is a five-class classification (classes C1 to C5) as shown in the defect map 420, first click the "Add Class" button 450. By pressing and making the specified settings, five defect classes with arbitrary labels are added and displayed in window 4 2 4. Although six classes are displayed in the window 4 24 in the figure, the labels 4 25 of the defect class C 1 b are displayed for later explanation and are not present. By the way, if you want to delete a defect class, you can select any defect class you want to delete in window 4 24 and remove it by pressing “Remove class” button 45 1.

1.3 Generating defect classifiers 120 (loop 304)

Next, a step of incorporating a classification criterion specified by the user (here, the five-class classification) as a classification criterion inside the system (for example, the processing terminal devices 102, 108) shown in FIG. 1 will be described. To this end, the above five steps (305 to 309) are performed as necessary, and the results are evaluated as needed to form the entire defect classifier 120. One of these processes Part or all can be automatically determined by learning, etc., but at the beginning of learning, it is effective to determine the items that can be set by humans as much as possible and to reduce the burden on system learning. Next, each of the five steps (305 to 309) will be described in detail.

1.3.1 Designation of decision tree for defect classification (Step 3 05)

The method of designating the defect classification decision tree in step 105 will be described with reference to FIG.

FIG. 5 is a diagram showing an embodiment of a setting window of a decision tree structure developed hierarchically and a teaching window of a defect sample according to the present invention.

In the present embodiment, the decision tree indicates a branching procedure for achieving the final classification of the defect class specified in the window 424 shown in FIG. Branch element ". Pressing the “Specify teaching sample / Class / Branch configuration” button 4 52 in Fig. 4 displays the window 500 shown in Fig. 5. Window 500 can be displayed and operated at the same time as window 400. Also, windows 400 and 500 may be displayed in the same window. At the stage when the defect class is specified in the window 424, the class elements 503 corresponding to the number of the defect classes constituting the classification decision tree are created in the window 502. In addition, a branch element 504 constituting a classification decision tree is prepared by default. Using these, the structure of the decision tree is specified in the window 501. The window # 501 in FIG. 5 shows an example of a completed decision tree, which is created through the procedures shown in FIGS. 6 (a) to (c).

FIG. 6 is a view showing an embodiment of a window for explaining the setting procedure of the hierarchically expanded decision tree structure shown in FIG. 5, and FIG. 6 (a) shows a defect class in the first branch. Fig. 6 (b) shows the case where the next branch is set in the first branch, and Fig. 6 (c) shows the case where the defect class is set in the second branch. Is shown.

First, as shown in Fig. 6 (a), the classification start point 601 and the first branch B1 (602) are displayed by default. If it is desired to branch the defect class C1 in the first branch B1, as shown in FIG. 5, a class element 518 having the label of the defect class C1 is transferred from the window 502 to the branch B1 (602). Drag, and drop. At this point, a copy 605 of the class element 518 is displayed below the branch B1 (602), as shown in FIG. 6 (b). Next, when it is desired to provide a branch in parallel with the class element 605, the branch element 504 is dragged, dropped, and dropped from the window 502 shown in FIG. 5 to the branch B1 (602). At this point, as shown in FIG. 6 (c), a branch element B2 (603), which is a copy of the branch element 504, is displayed below the branch B1 (602).

The branch elements are automatically or manually marked with a serial ID each time they are copied, for example, branch Bl, B2,…. Further, when it is desired to branch the defect class C4 at the branch B2, the class element 519 having the label of the defect class C4 is similarly dragged, dropped, and dropped to the branch B2 (603). The work is performed. The decision tree consisting of a combination of class elements and branch elements can be arbitrarily constructed according to the following three conditions.

(1) There is only one branch element in the layer immediately below the start point 601 of the classification.

(2) Any number of class elements and branch elements may be added immediately below any branch element.

(3) Nothing can be added immediately below the class element

(4) Any number of the same class element may be added immediately below different branch elements.

Here, every time a branch element is added, a window showing a list of branches 分岐 A branch label has been added to 426 (see Fig. 4), and to the right of each branch, each defect class or branch that branches from that branch is displayed with a label, such as color coding, so that it can be identified. Is done. As an example, the branch element B 2 (603) branches into three classes of defect class C 4 (606), 〇110 (607), and branch 83 (604). In statement 427, the label 448 of branch B2 and the frames and labels 449 corresponding to the three defect classes or branches that branch off from branch B2 are displayed.

As shown in FIG. 5, for example, with respect to the aforementioned combination condition (4), for example, the class elements 605 and 607 of the same defect classes C1 and C1b are respectively placed immediately below the branch elements B1 and B2. It is possible to attach. In this case, the defect classification results are later integrated. In order to make a distinction when multiple same class elements are created, a serial ID is automatically or manually attached each time it is copied, for example, a defect class Cl, Clb, CIc,…. Further, each time the same class element is copied, the label of the defect class copied in the window 425 is added like 425.

As described above, the structure of the decision tree that hierarchically expands and classifies defects is as follows: branch B 1 (602) branches to defect class C 1 and the remaining defect classes, and branch B 2 (603) Hierarchical, that is, branching to defect class C4, defect class C1b, and the remaining defect classes, and branching to defect class C2, defect class C3, and defect class C5 at branch B3 (604). By developing in stages, it is possible to easily, quickly, and surely set a classification standard that meets the user's requirements.

1. 3. 2 Defect sample teaching (Step 306)

The method of teaching the defect sample in step 306 will be described. First, the window 500 shown in FIG. 5 is displayed. Missing in window 424 in Fig. 4 At the stage of specifying the defect class, a number of frames corresponding to the defect class are created in the window 505. In FIG. 5, six frames of the defect class are displayed. However, in the window 501 where the decision tree is generated, until the copy C 1 b of the defect class C 1 is created, the frame 508 is displayed. not exist. The teaching of images to each defect class can be done by selecting one or more defect samples in the defect distribution map (408 or 409) or the image display window (402 or 403). If the selected defect sample is to be taught, for example, as the defect class C1, the image of the defect sample is dragged, dropped, and dropped into the frame 506 corresponding to the defect class C1. Teaching to other defect classes can also be performed by similarly sending the image group of the defect sample group within the frame of each defect class, and the presence / absence of teaching and the number of sheets need to be unified between defect classes There is no. Also, images of the same defect sample can be taught in a plurality of defect class frames. Also, even if the images of the teaching samples from the heterogeneous defect inspection equipment are the same defect locations, they can be taught as different teaching samples. Can be displayed). As described above, each time the defect class element is copied into the window 501, a frame for the defect class copied in the window 505 is newly created. For example, there are frames 506 and 508 for the class elements 605 and 607 of the defect class C1 and the copy of the defect class C1b. It is possible to teach the same defect sample group or different defect sample groups to these frames.

1.3.3 Evaluation of Degree of Separation of Defect Attribute Distribution (Step 307)

Visualize or quantify the degree of separation of attributes in the method of determining the classification rules (consisting of class elements and branch elements created according to the above rules) for each branch of a decision tree that is expanded hierarchically Talk about the method Step 307). However, it is not necessary for the entire decision tree to be completed at the time of this work, and it is not necessary to teach all the defective samples.Also, after this work, the structure of the decision tree and the teaching pattern of the defective samples are changed. In this example, the decision tree structure and the teaching of defect samples are all completed as shown in window 501 in FIG. 5. explain. First, a branch to which a classification rule is to be assigned is selected from a window 426 displaying a branch list shown in FIG. Here, as an example, if branch B 2 (4 27) is selected, three defect classes and branches (C 4, C 4) separated at branch B 2 in the defect attribute list window ゥ 4 5 4 lb, B 3) (The defect classes and branches are collectively referred to as categories.) The distribution of various attributes corresponding to the three defect classes and branches is displayed by being distinguished by a method such as color coding for each of the three defect classes and branches. 455, 46 ◦ etc.). In Fig. 4, instead of being color-coded, they are shown in white, dots, and diagonal lines. Naturally, for example, the feature amounts of the defect samples included in the defect distribution map 408 and the defect distribution map 409 for generating the defect classifier 120 are calculated in step 203 shown in FIG. It is assumed that Therefore, in the branch B2, the image of the teaching defect sample in the defect class C4 is indicated in the frame of 514, and the image of the teaching defect sample in the defect class C1b is indicated in the frame of 508. At the stage of teaching within, for example, the feature amounts of these taught defect samples are calculated, so the taught defect samples of two defect classes are obtained from the feature amount distribution of the entire defect sample excluding the defect class C1, for example. It can be displayed separately.

Here, three defect sample groups corresponding to the three defect classes and branches (three categories), which are displayed in different colors, are “defect sample groups taught as defect class C 4 5 5” and “defect classes”. Defective sample group 509 taught as class C 1b ”,“ Defect class C 2 under branch B3, Each of the defect sample groups 511, 513, and 517 taught as C3 and C5 '', and the distribution of these various attributes is displayed for each defect inspection apparatus from which each taught sample image was acquired. It is displayed in a separate window (455, 456) (it is also possible to combine and display). In the list of various attribute distributions displayed in the window 4 5 4, image features (color (texture, etc.) and size (indicated by area and length) of defect obtained from various defect images, shape (foreign matter) Not only quantified features such as shape, scratch shape, etc.) and the positional relationship of defects with respect to the wiring pattern (the positional relationship indicated by fatal defects such as short-circuit defects and disconnection defects) Defect classification results (results of any defect sample being classified into any defect classification class), composition analysis results

(The result of what kind of composition of an arbitrary defect sample is combined), and the numerical value (for example, Frequency) of the distribution of the defect map, etc. are included. In the window 4 5 4, the “Add defect attribute” button 4 2 2 is pressed and the specified setting is made to add a prepared attribute or design and add a new attribute. Or it is possible. Also, it is possible to delete any attribute. Incidentally, as described above, the decision of the decision tree and the teaching of the defect sample need not be completed, and are reflected in various attribute distributions within the specified range. As shown in FIG. 4, the difference in attribute distribution between the defect class or the defect sample group taught in the lower layer of the defect class or branch can be understood for each defect class or branch (category) classified in the designated branch. The display makes it clear which attribute is effective for classification in the branch.

As an example of a display method of various attribute distributions, a histogram display such as 559, 460 can be considered. However, the present invention is not limited to such a display method. 2 By pressing 1, for example, a two-dimensional or three-dimensional distribution display method using a combination of arbitrary attributes In addition, it is possible to select a display method in which the degree of separation of the attribute for each defect class or branch is clear to the user. Specific examples of the two-dimensional or three-dimensional distribution display method will be described later. In addition, by selecting an arbitrary defect sample, it has a function of displaying where the attribute in the defect sample exists in the entire attribute distribution and a function of displaying its numerical value _(in addition to the fact that the user As one of the indices for judging, the function of quantifying the degree of separation and displaying the quantified value of the degree of separation for each attribute, for example, 457, 458, etc., is provided. Examples of the method of numerical conversion include, for example, deviation and dispersion of the average value of the attribute distribution between the defect classes.

1.3.4 Selecting a classification rule (Step 3 08)

Next, in step 308, a method of generating a classification rule assigned to each branch will be described. First, an arbitrary branch to which a classification rule is to be assigned is designated from within the window 4 26. Here, as an example, the branch B 2 (4 2 7) is specified, and the classification rules inside the system shown in FIG. 1 for realizing the classification of the defect classes C 4, C lb, and the branch B 3 branching in the branch B 2 are shown. Describe how to decide. (At the stage where the branch B 2 (4 2 7) is specified, the attribute distribution list in the window 454 as described above includes the attribute of the defect class C 4, C 1 b, and the branch B 3. The distribution is displayed in different colors.) Next, when the "Classification rule designation" button 453 is pressed, the classification rule generation window 700 shown in Fig. 7 is displayed.

FIG. 7 is a diagram showing an embodiment of a classification rule generation window for each branch in the decision tree according to the present invention. Window 700 can be displayed and operated simultaneously with windows 400 and 500 (although they may be displayed in the same window). In this window 700, classification rules can be set according to the following two types of classifications: "rule-based classification", "learning-type classification (teaching-type classification)", and combinations thereof. If the designer has knowledge of the decision tree configuration, how to combine rule-based classification and learning-type classification, and the parameters of the rule-based and learning-type classifiers, if the designer has knowledge, it can be built-in or manual. The above setting items can be determined with. If it is not possible to assume that there is knowledge, it is conceivable to determine the above setting items by learning. However, in a situation where a sufficient number of teaching data cannot be ensured, such as when starting up the system, there is a risk that the above-mentioned over-learning occurs and the performance is reduced. Therefore, the distribution of defect images and attributes obtained from each defect inspection device is displayed, and based on such information, the setting items are determined as much as possible, and the degree of freedom of the items that must be determined during learning is reduced. It is effective.

1.3.4.1 Rule-based classification

Rule-based classification is a classification method that generates a classification rule using a combination of preset conditional expressions (consisting of items such as attributes, relationships, and thresholds (boundary lines or boundary surfaces)). First, when the rule-based classification is adopted, a check box 701 is checked. Next, an example of a method of specifying a conditional expression in the rule-based classification will be described. First, press the “Add Condition” button 70 2 to add at least one conditional expression. For each conditional expression, three items, "attribute (709)," "relation (710)," and "threshold (711)" are specified. In the figure, as an example, four conditional expressions 1 to 4 (705 to 708) are generated. As described above, a plurality of attributes are calculated for each defect sample, and the user selects an attribute having a high degree of separation effective for defect classification and incorporates the attribute into the conditional expression. However, an attribute with a high degree of separation for the teaching sample is not always an effective attribute for classification, and this judgment is left to the user. It is also possible to modify the rules by subsequent additional learning.

The degree of separation is based on the various attribute distributions displayed in window 4 54 as described above. This can be determined based on a reference (histogram, two-dimensional or three-dimensional display) or a numerical value of the degree of separation (described in 457, 458, etc.). For example, first, "attribute (709)" is selected using a pull-down menu. It is possible to select multiple attributes.

(1) When only one attribute is selected in the selection of “attribute (709)”, the remaining “relation (710)” and “threshold (71 1)” are set. For "Relation (7 10)", select the inequality sign (>,, =, ≤, <) using, for example, a pull-down menu. For "Threshold (7 11 1)", enter a numerical value using a keyboard (not shown). Pressing the “mouse input” button (712) on the right side of each condition column sets the “threshold (71 1)” intuitively using a mouse or sunset screen on another screen. It is possible to set. This will be described with reference to FIG.

FIG. 8 is a diagram showing an embodiment of a multidimensional graph display of various attributes according to the present invention and a method of designating a constraint condition in the graph. For example, in the graph 800 of the frequency of the attribute f1 shown in FIG. 8 (a), the threshold Th. 1 for the attribute f1 is obtained by moving the boundary 801 on the histogram horizontally with the mouse. (802) is determined and assigned to the “threshold (711)” column shown in FIG. Also, as shown in Fig. 8 (b), by simultaneously drawing the attributes f1 and f2 on a two-dimensional graph, it is possible to grasp the overall tendency, The threshold Th. 2 (807) can be set by moving. By the way, in this example, linear separation of 2 classes (in Fig. 8 (b), attribute distributions of samples belonging to each class are indicated by circles and triangles) is possible from three conditional expressions with boundary lines 804 to 806. This is a simple example.

(2) If two or more attributes are selected in “Attribute (709)” selection, fill the remaining “Relation (710)” and “Threshold (71 1)” boxes Since setting is difficult in such a format, leave these blank and press the “mouse input” button (712) to make settings on another screen.

For example, when two attributes are selected, a two-dimensional graph of the selected attributes 1 and 2 may be displayed as shown in Fig. 8 (c). Multiple boundaries can be specified, assuming that you want to classify the two classes displayed as white circles and black triangles. For each boundary line, for example, a straight line (808, 810, 812) is determined by designating two points on a two-dimensional graph, and an arrow () is used to specify the direction of which area to be set by the straight line. 809, 811, 813). A boundary line can be specified as a straight line or a line segment. The final condition specifies AND (logical product) or OR (logical sum) of conditions specified by multiple boundaries. In some cases, an exceptional defect sample (for example, defect sample 814) is included in the teaching sample due to the attribute distribution, and it is determined by the user whether or not to include these in the boundary. FIG. 8 (c) shows an example in which the defect sample 814 is excluded from the black triangle class. Also, as shown in Fig. 8 (d), it is possible to draw a free curve (815) with the mouse and specify the direction by using the arrow (816) to specify the condition.

Ό o

When three attributes are selected, a three-dimensional graph of attributes f1, f2, and f3 may be displayed as shown in Fig. 8 (e). Assuming a situation where it is desired to classify the two classes indicated by circles and triangles, the attribute space can be divided using a surface equation such as a spline surface or multiple identification surfaces approximated by a set of planar patches. . The three-dimensional attribute space displayed on the two-dimensional screen can be displayed from different viewpoints. In addition, it has a function to color-code and display each area divided by the identification surface to help the user understand. In Fig. 8 (e), each defect sample is color-coded into white and black according to which of the two regions divided by the identification surface. That is, the book In the example, if the round defect sample is colored white and the triangular defect sample is colored black, it can be said that the surface is a good identification surface. Next, an example of a method for adjusting the identification surface will be described. Since the defect sample 819 belongs to the triangular defect class, it is desirable to be colored black, but in Fig. 8 (e), it is colored white. By moving this to the opposite side of the identification surface and coloring it in black, any point 820 on the surface is moved (in some cases, any control point that does not exist on the surface is moved), and the identification surface is defective. Deform locally over sample 819. In FIG. 8 (f), the defect sample 819 is colored black. By repeatedly performing such adjustment, a good discrimination surface can be generated. Also, the degree of freedom of the identification surface can be set arbitrarily. -Next, combine the conditions 1 to 4 (705 to 708) specified by the above procedure to set the final conditional expression belonging to each defect class. The combination of conditions is determined by using logical expressions (AND (*), OR (ten), NOT (not), XOR (χ ο r)). If you want to take the value, write “1 * 3” in the box 713. In addition, it has a function of calculating and displaying a boundary line-identification surface candidate calculated by the processing inside the system as a reference value, and the user can make detailed adjustment using this as an initial value. The information of each set border and identification surface is stored internally, and can be called up and modified later.

1. 3. 4. 2 Learning type classification (teaching type classification)

Learning type classification is a classification method that basically generates classification rules by teaching. The learning-based classification method is effective for attributes whose conditions are difficult to set, such as the rule-based classification, even if there are attributes considered to be effective for the classification. When adopting this type of classification rule, check the check box 716 in FIG. Next, in a window 717, a plurality of attributes used for the learning engine are selected. Regarding attribute selection, Instead of manual selection, it is also possible to select automatically by a method such as performing weighting according to the validity or the like for each attribute by learning. However, such automatic selection of attributes may cause over-learning if the number of training samples is not sufficient, and may have an attribute distribution with a high degree of separation only for training samples. For some reasons, it has a function that can be combined with manual attribute selection and learning. Next, select the engine to be used for classification, such as the maximum likelihood estimation method or the K-NN method, from the pull-down menu of 718. There is also an automatic selection mode in this menu, which has a function to automatically select an appropriate engine according to the number of learning samples. This engine is basically applied when the conditional engine is not used or when the defect class is not determined in the conditional engine, but two branches in the window 501 in Fig. 5 are used. It is also possible to have a configuration in which the upper level is a learning type and the lower level is a conditional type.

The various attributes used in the display of the attribute distribution, the decision tree, the generation of the classification rules, and the like are: orthogonalization processing of the attributes by principal component analysis; Processing consisting of at least one combination of the compression processing of the attribute dimension by using only one or the rearrangement processing of the attribute distribution in the attribute space using the kernel function etc. (for defect samples belonging to different defect classes) It is possible to perform such processing that the attribute has a high degree of separation in the attribute space). The attribute redesigned in this way is added as a new attribute to the window 454 in FIG. It has functions that can be used as well as attributes. The combination of the above three processes has the advantage that defect classification can be performed using a simpler and clearer identification surface, but generally the attributes are difficult to understand physically. In the display function of the defect attribute distribution, the decision tree, and the determination method of the classification rule in the present invention, to a certain extent, there is no knowledge about the physical meaning of the attribute. It is possible to generate a defect classifier 1 20.

1.3.5 Classification result evaluation (Step 309)

The method of evaluating the generated defect classifier according to the present invention in step 309 will be described. This evaluation can be performed even if the decision tree for classification is not completely completed. After setting the classification rule for an arbitrary branch, by pressing the “re-classification” button 4 23 in FIG. 4, the defect classification using that branch is displayed on the map 4 20. If this result is not good, the teaching sample, decision tree structure, and classification rules are modified as appropriate, and if the result is good, the remaining classification rules are specified as in the loop in Fig. 3. The entirety of the defect classifier 12 according to the present invention is determined while performing 304 multiple times.

When all decision trees and classification rules have been determined, condition 310 is satisfied, and the process ends. Further, which defect inspection apparatus to use the generated defect classifier 120 according to the present invention is designated by check boxes 414 and 415 shown in FIG. In this embodiment, the setting is applied to the defect review device B.

[Second embodiment] (Generation of defect classifier 120 according to the present invention in defect review apparatus)

A second embodiment will be described. In the first embodiment, the method of generating the defect classifier 120 according to the present invention in one defect inspection apparatus has been described. The problem is what kind of classification class is assigned to each inspection device, and how to generate the defect classifier 120 according to the present invention that realizes the classification. For example, in a combination inspection using a defect detection device and a defect review device, if a defect classification class that could not be completely classified by the defect detection device can be classified in detail by the defect review device becomes possible, a defect Review equipment It is considered that the number of defect classification classes at the time of classification can be narrowed down, and effective learning of the defect classifier can be performed. Conventionally, however, the defect classification standards for both defect inspection systems have been set individually, and hierarchical defect classification classes based on unified classification standards have not been allocated. Therefore, the present invention provides a method of generating a defect classification class and a defect classifier for performing hierarchical defect classification according to the inspection order.

The method for generating a defect classification class and a defect classifier in the present invention is also effective for controlling a review sampling plan. For example, if a defect sample classified into an arbitrary defect classification class by the defect detection device is later difficult to obtain useful information on defect classification even after analysis by the defect review device, the review sample A method of reducing the number is conceivable.

In the present embodiment, assuming that defect detection has already been performed in any defect detection device, the defect review device integrates the inspection information obtained from the defect detection device and the defect review device. This paper describes how to generate defect classification classes and defect classifiers to achieve effective and detailed automatic defect classification that satisfies the user's classification requirements. The defect samples inspected by the defect review device are a set sampled from the defect samples inspected by the defect detection device. Therefore, for all defect samples for which defect classification is performed by the defect review device, both inspection information obtained from the defect detection device and the defect review device can be used.

2.1 Processing flow

In the following description, the analysis method for each combination of the defect detection device and the defect review device will be described in particular. However, the same analysis can be performed for any combination of three or more defect inspection devices. Fifth embodiment Described in). Similar analysis can be performed for a defect sample common to a defect detector and a defect reviewer other than the combination. The defect distribution maps 91 to 903 in FIG. 9 show, as an example, the distribution of defect samples on the defect distribution map and the defect classification results at each processing stage.

(1) First, after processing in a predetermined processing step of semiconductor device manufacturing, inspection by a defect detection device is performed. The inspection information obtained by the inspection is sent to the data server 107 or the processing terminal device 108 as needed. The defect distribution map 901 shows, for example, the distribution of defect samples in the defect detection device and the result of coarse defect classification (before adjustment) by defect classification in the defect detection device. 2. One defect sample is classified into three defect classification classes C a1 to C a3. However, it is not essential to perform defect classification in this step.

(2) The defect sample group detected in step 901 is sampled for review inspection as necessary (this is called review sampling).

(3) Inspect the defect sample group subjected to the review sampling with a defect review device. However, the processing steps of the semiconductor device manufacturing to be reviewed here need not be the same as the processing steps inspected in step 91. The inspection information obtained here is sent to the data server, '107 or the processing terminal device 108 as necessary.

(4) Based on the inspection information obtained by the defect detection device and the defect review device obtained in (1) and (2), respectively, the defect classification class of the defect review device is determined and classified into the defect classification class. Generate a classifier. By using the inspection information of both the defect detection and review devices, it is possible to perform defect classification that was difficult to classify using the defect review device alone. The defect distribution map 902 is inspected by a defect review device as an example. This is a detailed classification result (before adjustment) in which defects are automatically classified using only the defect position where the defect was performed and the inspection information obtained from the defect review device. 2 This shows how defect samples that have been reviewed and sampled from 1 to 9 points are classified into four defect classification classes Cb1 to Cb4. Here, the defect samples da 1 and da 2 on the defect distribution map 90 1 are classified into different defect classification classes C a 2 and C a 3, whereas the defect samples da 1 and da 2 on the defect distribution map 90 2 The defect samples db1 and db2 corresponding to the defect samples da1 and da2 are both classified into the same defect classification class Cb3. Conversely, in some cases, the corresponding defect samples are classified into the same classification class in the defect distribution map 901, and are classified into different classification classes in the defect distribution map 902. If defect classification is performed using the inspection information of both inspection devices for such inconsistency in the classification results between the defect inspection devices, the defect samples can be classified as the same classification class or finely classified. It is possible. The defect distribution map 903 combines the inspection information obtained in 901 and 902, and is an example of the detailed classification result (after adjustment) of the defect classification optimized according to the classification requirements of the user It is. For example, defect classifiers for subclassifying defect samples da 1 (db 1) and da 2 (db 2) are used (defect sample db 3 is defect class C b5, defect sample db 4 is defect class Class Cb3), showing how defects were classified into five defect classification classes Cbl to Cb5. The method of generating the defect classification class and the defect classifier in the second embodiment can be performed in the same manner as the procedure in the first embodiment. However, in the second embodiment, the defect according to the present invention is based on the attribute information obtained from both the defect detection and review devices displayed in windows 455 and 456 shown in FIG. A classifier can be generated.

According to the above procedure, the setting of the defect classifier according to the present invention in the defect review device is completed. Once a defect classifier has been created, And subsequent wafer inspection is continued. However, it is possible to continuously change or additionally learn the defect classifier based on the inspection information obtained thereafter.

2.2 Change of image processing procedure or adjustment of image processing parameters When integrating and using inspection information between defect inspection equipment, consistency between the inspection information may become a problem. The case where the matching is not achieved is, for example, the recognition result of the wiring region in the same defect sample, the positional relationship between the defective region and the wiring region (isolated single line / overlay judgment), the size and height of the defect, Or, there is a case where the defect attributes such as the vertical relationship with the film formation of the defect (determination on the film under the Z film) differ among a plurality of defect inspection apparatuses. Hereinafter, an embodiment of the matching method will be described.

FIG. 13 is a diagram showing image processing results of a defect image of the same defect sample in the defect detection device (ordinary defect inspection device) A and the defect review device B. Inspection images 1301 and 1302 indicate a reference image and a defect image captured by the defect detection device A, respectively. Also, arbitrary image processing A (1303) and B (1304) are performed on the reference image 1301 and the defect image 1302, and the binarized images of the wiring area are the binary images 1305 and 1306, respectively. However, in the binarized image 1306, the defect area (indicated by a white circle in the figure) is also binarized and displayed. Reference images and defect images in the defect review device B corresponding to 1301 to 1306, image processing C and D, and binarized images are 1307 to 1312, respectively. Here, as an example, in the binary images 1305 and 1306, wiring recognition has failed, and an example has been shown in which the central wiring is not extracted as a binarized region. The actual defect area is a highly fatal defect existing over two wirings, and is determined to be an isolated defect from the binary image 1312 in the defect review apparatus B, but is determined in the defect detection apparatus A to be an isolated defect. From the binary image 1306, Has been refused. It is necessary to change the image processing procedure or adjust the image processing parameters so as to match such differences in defect attributes.

Therefore, from the binary image 1311 obtained by the defect review device B, processing such as magnification change, distortion correction, or brightness correction in the case of a grayscale image is performed, and the wiring binary image in the defect detection device A is processed. The teaching pattern 1 3 1 3 is generated as the correct pattern of. Next, the image processing procedure in the image processing 1303 is changed or the image processing parameters are adjusted so that a processing result that matches or is similar to the teaching pattern 1313 is obtained in the binary image 1305. U. The same applies to a defect image, but in the defect image, not only the wiring area but also the size of the defect area may be matched. In the first place, the reference image 1301 and the defect image 1302 in the defect detection device A given as an example are different from the reference image 1307 and the defect image 13008 in the defect review device B in contrast and contrast. The resolution was inferior, and setting of the image processing parameters was a difficult example. Based on the above-described procedure, not only matching of inspection information among a plurality of inspection apparatuses but also setting of various image processing parameters which are difficult to set can be performed.

In the present embodiment, the defect review device B was selected as the defect inspection device for acquiring the inspection images 1311 and 1312 that generate the teaching patterns 1313. The selection of which defect inspection apparatus to generate the teaching pattern can be determined by the user from the simultaneous review screen of various inspection images, or can be automated by setting an arbitrary rule. This processing can be performed in all combinations of defect inspection devices.

[Third Embodiment] (Generation of a defect classifier 120 according to the present invention in a defect detection device (ordinary defect inspection device))

In the third embodiment, based on the result of the detailed inspection by the defect review device, the defect review device effectively performs defect classification. The method of determining the defect classification class and defect classifier in the defect detection device, and the review sampling method are described below. In order for defect classification in a defect review device to be performed effectively, it is desirable that the defect classification class in the defect detection device be close to the defect classification standard in the defect review device. If it is possible to perform a hierarchical classification, such as performing a detailed classification with a defect review device on a defect classification class that could not be classified by a defect detection device, a defect detection device that is not subdivided in a review inspection device The number of review samples can be reduced for defect samples classified into the defect classification class. In the defect review system, the number of defect classification classes can be reduced, and effective learning of the defect classifier becomes possible. The third embodiment is based on the premise that the defect classification class in the defect review device is known. However, it is not essential that the defect classifier in the defect review apparatus described in the second embodiment is customized. That is, the determination of the defect classifier of the defect detection apparatus according to the third embodiment corresponds to the defect distribution maps 93 to 95 in FIG. 3 can be carried out after the customization such as the defect distribution map 91-903 or without the customization. However, when the defect classifier in the defect review device described in the second embodiment is customized, the defect classifier in the defect detection device in the third embodiment is customized. It is expected that the defect classification class Cb in the defect review apparatus can be set to be a subset of the defect classification class C a in the defect detection apparatus or a set close to the subset by the evening maze. The following description is based on the assumption that the defect review class was implemented following the customization of defect classifiers and defect classifiers. In the following description, as in the second embodiment, the analysis method using a single defect detection device and defect review device will be particularly described. Similar analysis is possible for a combination of defect inspection devices.

3.1 Determination of defect classification class

The defect review device so that the defect classification class in the defect detection device (first defect inspection device) and the defect detailed classification in the defect review inspection device (second defect inspection device) performed thereafter are effectively performed. Set to be a class similar to the defect classification class in. Here, the defect classification class C b in the defect review device (second defect inspection device) is a subset of the defect classification class C a in the defect detection device (first defect inspection device), or a similar subset. Set to be classified. As an example, as shown in a defect distribution map 903, assuming that defect classification classes C bl to C b 5 in a defect review device are given, a defect detection device can be obtained based on the defect classification classes C bl to C b 5 A teaching pattern of a defect classification class in the defect detection device is created as a defect distribution map 904 using only the inspection information. Next, a defect classifier in the defect detection device for realizing the classification into the defect classification classes similar to the teaching pattern is generated. Here, the difference from the defect classifier in the defect review device in the second embodiment is that only the inspection information that can be used is obtained from the defect detection device. The point is that the obtained test information cannot be used. That is, the inspection information that can be used in the defect detection device of interest is only the inspection information obtained before the defect detection device of interest in the inspection order in the actual inspection. The method of changing the defect classifier can be the same as the method of setting the defect classifier described in the first embodiment. As an example, a defect classifier that classifies the defect classification class in the defect distribution map 904, which is a teaching classification pattern, as much as possible in a defect detection device is generated, and defect classification is performed using the generated defect classifier. The result of the coarse classification (after adjustment) is the defect distribution map 905. In this embodiment, the defect distribution map This is an example in which the defect classification classes C a2 and C a5 could not be classified. (The defect classification classes C a2 and C a5 in the defect distribution map 904 are the defects in the defect distribution map 905.) Classification class C a2) and other defect classification classes C a1, C a3, and C a4 are subjected to defect classification similar to the teaching classification pattern 904. According to the above procedure, the setting of the defect classifier in the defect detection device is completed. Once a classifier is created, subsequent wafer inspections are continued using the defect classifier. However, it is possible to continuously change the defect classifier based on the inspection information obtained thereafter ₍ here, the defect detection maps shown in the defect distribution maps 91 to 95 described above, Fig. 11 summarizes the relationship between the defect classification classes for the procedure for determining each defect classification class in both review devices. The vertical column shows the defect classification class in the defect detection system, the horizontal column shows the defect classification class in the defect review system, and the figures in the columns indicate the defect detection class for the defect sample in the defect review system. The number of classifications classified into each classification class in both the review and review equipment is shown.The defect classification class name in the item column corresponds to Fig. 9.

(1) Table 111 shows the defect classification class and the inclusion relationship between the defect classification classes in both defect inspection systems before the defect classifier is adjusted (the defect distribution map 901 shown in Fig. 9 and the defect distribution class). Corresponding to the relationship with the distribution map 902). For example, the area surrounded by the frame 1104 indicates that the sample classified into the defect classification class Cb3 by the defect review device is the defect classification class C a2 or C a3 by the defect detection device. Indicates that it was classified as either. First, in the first adjustment of the defect classification class in the defect review device, for example, the defect classification class Cb in the defect review device is a subset of the defect classification class Ca in the defect detection device. As shown by 1104, one defect classification class Cb3 in the defect review equipment is missing. The defect classifier is adjusted and the defect classification class is subdivided so that two or more defect classification classes C a 2 and C a 3 in the defect detection device do not correspond. At this time, an unnecessary defect classification class may be deleted, a new defect class may be added, or a defect class may be rearranged according to a user's classification request.

② Table 1102 shows the inclusion relationship between the defect classification class in the defect review device after adjustment obtained by the above adjustment and the defect classification class in the defect detection device before adjustment (shown in Fig. 9). This corresponds to the relationship between the defect distribution map 901 and the defect distribution map 903). When compared with Table 1101, the defect classification classes classified as defect classification class Cb3 in Table 1101 were Ca2 and Ca3. Since the defect classification class was only Ca3 in the above, for example, in response to a request to review many defects belonging to the defect classification class Ca3 in Table 1102, the defect classification class Ca2 Since the candidate is no longer a candidate of class C a3, the classification can be effectively realized by increasing the number of reviews of the defect sample classified into the defect classification class C a3.

Next, the defect classification class C a in the defect detection device is adjusted to be similar to the defect classification class C b in the defect review device. Specifically, the defect classification classes Cb2 and Cb5 and the defect classification classes Cb3 and Cb4 in the frames 1105 and 1106, respectively, should be subdivided in the defect detection device. Adjust the defect classifier to However, since the defect detection device cannot use the inspection information in the defect review device, there may be a situation where the reliability of the classification cannot be obtained with respect to the defect classification performance in the defect review device. In such a case, it is conceivable to perform processing such as not dare subclassing, or increase the number of review samples to confirm even if subclassifying.

③ Table 1103 shows the inclusion relationship of defect samples belonging to the defect classification class in both defect inspection devices after the adjustment obtained by the above adjustment (No. 9). This corresponds to the relationship between the defect distribution map 903 and the defect distribution map 905 shown in the figure). When compared with Table 1102, the defect classification class C a3 in Table 1102 is divided into defect classification classes C a3 and C a4, and the defect classification classes C b3 and C b4 are classified. It is possible. In the present embodiment, there is no change in the relationship between the defect classification classes in the frame 1105 even in the frame 1107. This is considered to be the case where it is difficult to classify only from the inspection information obtained from the defect detection device in the first place, or the case where the classification is performed intentionally due to the reliability of the classification as described above. As a result, focusing only on the correspondence of the defect classification classes, the final detailed classification is possible if the defect detector performs the review inspection only on the defect samples classified into the defect classification class C a 2. In practice, a processing method is conceivable in which the ratio of review sampling is increased in defect classification class C a 2.

In determining the defect classification class, it is effective to clearly show the user the relationship between the defect classification classes in a plurality of defect inspection apparatuses. FIG. 11 is an example of a display method showing the relationship between the defect classification classes. By displaying the number of classifications for each defect classification class, the degree of overlap between the classification classes can be known.

3.2 Review sampling method

Next, the review sampling method is described. As described above, if the defect classification class C b in the defect review device is set to be a subset of the defect classification class C a in the defect detection device, or a similar set, the defect detection device performs defect detection and review. The need for review inspection is low for defect samples classified into the defect classification class common to both devices. Conversely, it is highly necessary to conduct a review inspection on defect samples classified into a defect classification class that is further classified into a detailed class in the review inspection. That is, according to the necessity of the review inspection It is conceivable to control the number of samplings. The defect distribution maps 903 and 905 are examples for explanation, and the total number of defect samples is extremely small, so if only the defect point da3 in the defect distribution map 905 is sampled for review, the entire defect detailed classification will be performed. This is an example where the image can be grasped, but such a case is not realistic. In practice, the number of defects is very large, and the image quality of the defect detector is generally inferior to that of the review inspection. Therefore, it is not expected that the defect detector will be classified into a detailed defect classification class. In addition, the classification results of classified defect samples are not always reliable. Therefore, in the actual inspection, defect sampling maps classified into any of the defect classification classes were subjected to review sampling at several points, and the defect distribution map was difficult to classify in detail with respect to the teaching classification pattern. For defect classification classes such as a2 (including multiple defect classification classes Cb2 and Cb5 in defect review equipment), a method of sampling more than other defect classification classes is used. Another method is to change the number of samples according to the reliability of the defect classification result.

As described above, the rate at which review sampling is performed can be determined for each defect classification class in the defect detection apparatus, for each defect sample, or in consideration of both. It is possible. In particular, in the judgment for each defect sample, it is effective to use the reliability of the defect classification result as a judgment factor. Regarding the reliability of the defect classification result, by defining the degree of belonging to each defect classification class, each defect sample is divided into:

(1) Defects classified into any defect classification class

(2) Boundary defects in which the degree of membership is balanced among arbitrary multiple defect classification classes (difficult to determine)

(3) Unknown defects that did not exist at the time of learning It becomes possible to classify into three. For (2) and (3) above, it is conceivable that the need for review inspection is large, and for (1), when the degree of belonging is low, the need for review inspection is large. As described above, by generating the defect classification class and the defect classifier 120 in the defect detection device and controlling the number of sampling points according to the defect classification class, the plurality of inspection devices have hierarchical consistency, and Effective defect inspection can be performed.

By the way, it may be difficult to set a defect classifier such that the defect classification class Cb in the defect review device is a subset of the defect classification class C a in the defect detection device. FIG. 12 is an example of this, and the classification classes (Cal to Ca3 and Cb1 to Cb4, respectively) in the defect distribution maps 1201 and 1202 by the defect inspection devices A and B do not have a subset relationship. However, even in such a case, for example, a defect sample classified into the classification class C a 3 by the defect inspection device A is classified into either the classification class Cb 2 or Cb 3 by the defect inspection device B. It is highly possible and has the effect of narrowing down the classification class candidates. Therefore, for defect samples classified into classification class Ca3 in the defect inspection device A, configure the defect classifier as a two-class problem of Cb2 or Cb3 instead of the three-class problem in the defect inspection device B. As a result, improvement in classification performance can be expected.

Further, in the second and third embodiments, the generation of the defect classification class and the defect classifier of the defect detection and defect review device is automatically generated to some extent based on the user's final classification request using learning or the like. It is possible. However, sufficient teaching samples cannot be obtained, especially at the start-up of the process, and there is a high risk that the defect classification classes and defect classifiers specialized in the unique properties of the teaching samples will be generated. . In the first embodiment, a detailed inspection information review and a defect classifier customizing method are used. This makes it possible to easily incorporate user requirements and knowledge regarding defect classification into the system, and to suppress peculiar classification rules. In addition, it is easy to change the classifier in the additional learning. For example, regarding the teaching sample, in a situation where a sample belonging to the defect class C a happens to have many round defects and a sample belonging to the defect class C b has many square defects, the decision tree is used. Even if a branch due to a difference in shape is set, such a classification rule can be canceled if the difference in shape is not an essential difference between the two classes. In addition, it is expected that during additional learning, by simultaneously observing changes in attribute distribution in multiple defect inspection devices, it is possible to observe a tendency for the degree of separation between defect classes of attributes that are not essentially different to deteriorate. You.

Next, regarding a defect sample group detected by the defect detection device, a method of estimating a detailed classification result of the entire defect sample group by performing a review inspection of a part of the defect sample group and performing detailed classification is described in detail. The method of using the ratio of the classification results is described. By performing a detailed inspection on a part of the defect sample groups selected by review sampling from the defect sample group, it is possible to grasp the ratio of the detailed classification result included in the part of the defect sample groups (defect detection When random sampling is performed for each defect classification class in the apparatus, the ratio of the detailed classification result for each defect classification class can be grasped.) For a defect sample for which a review inspection has not been performed, the inspection information obtained by the review inspection cannot be directly used, but the reliability of estimation can be improved in consideration of the ratio of the detailed classification results. For example, assume that the defect ratios of the defect classification classes C a and C b are a% and b%, respectively. The defect samples are sorted with respect to any attribute classified by the defect classification classes C a and C b, and the upper a% closer to C a in the boundary cases is classified as the defect classification class C a. [Fourth embodiment]

In the fourth embodiment, a description will be given of an inspection method in a case where respective defect sample groups in a plurality of defect inspection apparatuses sharing inspection information cannot partially refer to each other. If the defect inspection devices that share the inspection information are both defect detection devices, for example, if defect detection is performed independently in the combined inspection of the optical pattern inspection device and the SEM type pattern inspection device, Partial inconsistency can occur in the defect samples detected by both. In the present embodiment, the same analysis and classification as in the second or third embodiment can be performed on the portion where the defect sample is found or included. The following description will particularly describe an analysis method for a combination of two defect detection devices A and B. However, a similar analysis can be performed for a combination of three or more defect inspection devices. Similar analysis is also possible for combinations other than the defect detection devices if there is a mismatch in the defect samples. A method of interpolating the inspection information with the inspection information of another similar defect sample when the inspection information of the defect sample to be referred to is missing will be described. FIG. 12 shows, as an example, the positions where the inspections were performed in the two defect inspection apparatuses A and B, respectively, on the defect distribution maps 1201 and 1202. Here, consider the defect classification of the defect sample da1 on the defect distribution map 1201 inspected by the defect inspection apparatus A. However, inspection is not performed by the defect inspection apparatus B on the coordinates dbl on the defect distribution map 1222 corresponding to the defect sample da1. Therefore, each defect sample in the defect inspection apparatus B is clustered based on the spatial distribution of defects, the defect classification result, or various types of inspection information, and, for example, a class Cbl to Cb4 is obtained. If the degree of spatial belonging belonging to the class can be defined for arbitrary coordinates on the defect distribution map, inspection information common or similar to the class can be shared within the class. Noh is considered. For example, as the inspection information at the coordinate db1 on the defect distribution map 1222, the inspection information of the defect sample db2 considered to belong to the same class Cb1 as the coordinate db1 can be used. It is also possible to interpolate the inspection information from other defect sample groups belonging to the same class Cb1.However, the inspection may be performed due to errors in the cluster setting, errors in the inspection information of the reference destination, or fluctuations in the inspection information. There is a risk that the reliability of information interpolation may be impaired. Therefore, weighting is performed on the reliability of the class information, the degree of belonging to the class setting, or the reliability of the test information interpolated based on the reliability of the test information and parameters such as fluctuations, and the test information is obtained. It may be used when using.

[Fifth embodiment]

In the fifth embodiment, a description will be given of a combination inspection using three or more defect inspection apparatuses. In the second, third, and fourth embodiments, the combination inspection of two defect inspection apparatuses has been particularly described. However, the same analysis can be performed for the combination inspection using three or more defect inspection apparatuses. As an example

Figure 10 (a) shows the defect classification class and the procedure for setting the defect classifier (generation procedure) for the combined inspection of N defect inspection devices. As shown in FIG. 10 (a), the generation of the defect classifier 120 including the defect classification class is performed in order from the Nth defect inspection device to the first defect inspection device in reverse order of the inspection order. Will be This is because the defect classification is sequentially supplemented for defect samples that cannot be classified or whose classification is not reliable according to the inspection order. Therefore, the classification class in the (n + 1) th defect inspection device (second defect inspection device) in which inspection is performed after any n (0n ^ N) th is the defect inspection device in which nth inspection is performed. It is determined to be a subset of the classification class in (first defect inspection device) or a defect class similar thereto. That is, first, a classification class for the Nth defect inspection apparatus is generated, and thereafter, In order to limit the defect classes in the defect inspection apparatus to be inspected in advance, the classification class of the (N-1) to first defect inspection apparatus to be inspected before that is determined in order. In addition, the defect classifier in the arbitrary η-th defect inspection device is the defect inspection device of the 1st to (η-1) th defect inspection device and the η-th defect inspection device which have been inspected before that. The obtained inspection information 1002, which can be used as a criterion for defect classification, is selectively and comprehensively incorporated into the classification rule, thereby setting in the η-th defect inspection apparatus. A defect classifier capable of classifying the classification class is generated (step 1001). However, when creating defect classification classes and defect classifiers, it is possible to make decisions while referring to the inspection information obtained from all defect inspection equipment using the simultaneous review screen described in the first embodiment. It is. The defect classification class and defect classifier can be corrected iteratively as shown in Fig. 3, and when a defect class or defect classifier in any inspection equipment is generated, any defect class or defect classifier can be modified. It is also possible to change the classifier again. The process flow when classification is performed for actual data

Fig. 10 (b). As shown in FIG. 10 (b), the actual inspection is performed sequentially from the first defect inspection apparatus using the generated defect classifier 120, and if necessary, the third inspection is performed. Inspection by the next defect inspection device is performed while sampling a defect sample as in the above embodiment. However, Fig. 10

As an exceptional case in (a) and (b), after inspection with a plurality of defect inspection devices in any combination, defect classification is performed or sampling is performed using the inspection information obtained from the plurality of defect inspection devices. In such a case, the inspection information obtained from the plurality of defect inspection apparatuses can be used as a criterion of a defect classifier in the plurality of defect inspection apparatuses regardless of an inspection order. Industrial applicability As described above, according to the present invention, it is possible to easily perform customization of a defect classifier, which has been difficult in the past, in response to a user-specific classification request, and perform automatic defect classification that satisfies the user's determination criteria. Mechanism can be obtained. Further, according to the present invention, when the user clarifies his or her classification request, various inspection information is displayed on the almost simultaneous review screen so that the user can review the information at almost the same time, thereby giving a unified opinion. It becomes possible.

Further, according to the present invention, it is possible to realize a hierarchical defect classifier in combination inspection by a plurality of defect inspection devices, and to realize effective learning and effective review sampling of the defect classifier. Becomes

Claims

The scope of the claims

1. A method of generating a defect classifier for classifying the defect based on inspection information obtained by inspecting a defect generated in an object to be inspected by an arbitrary defect inspection apparatus,

An inspection information acquisition step of inspecting a defect sample group on an arbitrary sample by at least an arbitrary defect inspection device and acquiring sample inspection information;

A display step of displaying a state of a defect attribute distribution of a defect sample group on the arbitrary sample on the screen based on the sample inspection information obtained in the inspection information obtaining step on the screen; and the displayed defect attribute on the screen. A classification rule setting step of setting an individual classification rule for each branch element in a decision tree in which the classification class elements of the defect sample group are hierarchically expanded via the branch elements based on the distribution state. A method for generating a defect classifier, comprising setting steps.

2. In the displaying step, the defect attributes in the state of the defect attribute distribution include an image feature amount, a defect classification result, a defect coordinate, a composition analysis result, a construction history, an apparatus QC, and a defect position detected on a wafer. The method for generating a defect classifier according to claim 1, comprising information on distribution and a combination of at least one or two or more of the number of defects.

3. The method for generating a defect classifier according to claim 1, wherein, in the displaying step, the state of the defect attribute distribution is a graph showing at least a defect distribution map and a relationship between a feature amount and frequency.

4. In the displaying step, the state of the defect attribute distribution is a multidimensional graph in which a plurality of attributes determined by a group of teaching defect samples belonging to a category classified for each branch element in the decision tree are combined. 2. The method for generating a defect classifier according to claim 1, wherein:

5. A method of generating a defect classifier for classifying the defect based on inspection information obtained by inspecting a defect generated in an object to be inspected by at least one defect inspection apparatus,

An inspection information acquisition step of inspecting a defect sample group on an arbitrary sample with a plurality of defect inspection devices to acquire a plurality of sample inspection information;

A review step of displaying a plurality of sample inspection information acquired by the inspection information acquisition step almost simultaneously on a screen and browsing the same to determine a classification class element of the defect sample group;

A decision tree designating step for designating a decision tree that is hierarchically expanded through the branching elements based on the classification class element determined in the review step on the screen, and the examination information obtaining step on the screen And a classification rule setting step of setting an individual classification rule for each branch element in the designated decision tree based on at least one sample test information obtained by Method of generating defect classifiers.

6. The review step is characterized in that the plurality of sample inspection information viewed at substantially the same time is at least a plurality of defect distribution maps and a defect image of a defect sample designated on each of the defect distribution maps. 6. The method for generating a defect classifier according to claim 5, wherein:

7. The classification rule setting step is a multidimensional graph in which a plurality of attributes determined by a group of teaching defect samples belonging to a category classified for each branch element in the decision tree are combined as the one sample inspection information. 6. The method for generating a defect classifier according to claim 5, wherein:

8. In the inspection information acquiring step, the optional defect inspection apparatus is at least one of an optical or SEM type foreign substance inspection apparatus and a pattern inspection apparatus, a defect review apparatus, a scanning probe microscope, and an element analysis apparatus. Or two or more 2. The method for generating a defect classifier according to claim 1, wherein the method includes the above combination.

9. In the decision tree setting step, further, on the screen, the degree of attribute separation in an attribute space between defect sample groups belonging to each classification class separated for each branch element in the decision tree is visualized. 2. The method for generating a defect classifier according to claim 1, further comprising a step of evaluating a degree of separation of a defect attribute distribution to be evaluated.

10. The decision tree setting step further includes a teaching step of teaching and displaying a group of teaching defect samples belonging to a category classified for each branch element in the decision tree on the screen. A method for generating a defect classifier according to claim 1.

11. The defect classifier according to claim 1, wherein in the classification rule setting step, an individual classification rule for each branch element in the decision tree is a rule-based classification and / or a teaching type classification. Generation method.

12. In the classification rule setting step, if the classification rule of a certain branch element in the decision tree is rule-based classification, a condition sentence that is satisfied by a plurality of attributes for rule-based classification on the screen is set. The method according to claim 1, wherein the defect classifier is generated.

13. In the classification rule setting step, when a classification rule of a certain branch element in the decision tree is a teaching type classification, an attribute for performing the teaching type classification can be designated on the screen. How to generate the described defect classifier.

14. An inspection information acquisition step of inspecting a defect sample group on an arbitrary sample with a plurality of defect inspection devices to acquire a plurality of sample inspection information;

On the screen, a map showing a plurality of defect distribution states obtained from each of the plurality of sample inspection information obtained in the inspection information obtaining step is substantially displayed. A display step of displaying the defect images and the attributes of a plurality of defect samples specified on the map at the same time side by side at the same time.

15. The second defect inspection apparatus, which performs inspection after the inspection by the first defect inspection apparatus, inspects the defect generated in the inspection target for the defect, and classifies the defect based on the inspection information obtained. A method of generating a second defect classifier, wherein the second defect classifier is inspected by the first and second defect inspection devices for a defect sample group on an arbitrary sample. A method for generating a defect classifier characterized by being generated or changed using part or all of the information.

16. The first defect inspection device, which is inspected before inspection by the second defect inspection device, inspects the defect generated in the inspected object and classifies the defect based on the inspection information obtained. A method of generating a first defect classifier, wherein a defect classification class classified by a second defect classifier in the second defect inspection device is a defect classification classified by the first defect classifier. A method of generating a defect classifier, comprising generating or changing the first defect classifier so as to be a subset of a class or a classification similar thereto.

17. The method for generating a defect classifier according to claim 15, wherein the defect classifier is constituted by a decision tree in which a classification class element is expanded by a branch element.

18. The method for generating a defect classifier according to claim 16, wherein the defect classifier comprises a decision tree in which a classification class element is expanded by a branch element.

19. An inspection information acquisition step of inspecting a defect sample group on a sample by at least a combination of the first and second defect inspection devices to obtain first and second inspection information.

Referring to the first and second inspection information acquired in the inspection information acquiring step, the first or second defect information is determined according to the inspection order of the first and second defect inspection apparatuses. A defect classifier generating step of generating a defect classifier.

20. In the defect classifier generation step, when the first or second defect classifier is generated, the defect classification class in the first defect inspection apparatus for the defect sample group on the sample and the second 20. The method according to claim 19, further comprising a display step of displaying, in a diagram or a table, an inclusive relation between the defect class and the defect class in the defect inspection apparatus.

21. In referring to the inspection information in the defect classifier generation step, for the defect sample for which the inspection information has not been obtained, the inspection information is substituted or interpolated by the inspection information of another defect sample. 21. A method for generating a defect classifier according to claim 19.

22. The other defect samples are obtained from the spatial distribution of the defect sample group, or from each class obtained based on the defect classification result or various inspection information, and from the inspection information. 22. The method according to claim 21, wherein the selection is made based on the degree of attribution to a defect sample that has not been determined.

23. A first inspection information obtaining step of obtaining a first sample inspection information by inspecting a defect sample group on a sample to be inspected by a defect detection device;

A pre-adjustment coarse defect in which a defect sample group is classified into a pre-adjustment first defect classification class by a pre-adjustment coarse defect classifier based on the first sample inspection information acquired in the first inspection information acquisition step; A second inspection information acquisition step of inspecting a review defect sample group selected from the defect sample group on the sample by a defect review device to acquire second sample inspection information;

Based on the second sample inspection information acquired in the second inspection information acquiring step, the defect sample group for review is subjected to the pre-adjustment detailed defect classifier by the pre-adjustment detailed defect classifier. A pre-adjustment detailed defect classification step of classifying the defect into a second defect classification class; and the pre-adjustment detailed defect based on the first and second sample inspection information obtained in the first and second inspection information obtaining steps. A defect for teaching by adjusting the second defect classification class before adjustment classified in the classification step so as to correspond to the first defect classification class before adjustment classified in the coarse defect classification step before adjustment. A creation step for teaching to create a classification class;

The coarse defect classifier after adjustment is adjusted by adjusting the first defect classification class before adjustment classified in the coarse defect classification step before adjustment to be the defect classification class for teaching created in the creating step for teaching. Generating a coarse defect classifier for generating a coarse defect classifier.

24. An automatic defect classification method for inspecting an object to be inspected by a first defect inspection device, and then inspecting the object to be inspected by a second defect inspection device to classify defects occurring in the object to be inspected. ,

Based on the relationship between the defect classification class classified by the first defect classifier in the first defect inspection device and the defect classification class classified by the second defect classifier in the second defect inspection device. An inspection sample determination step for determining a sampling rate of an inspection sample for each defect classification class in the first defect inspection apparatus;

An inspection step of inspecting an inspection sample in accordance with the sampling ratio determined for each defect classification class in the inspection sample determination step with a second defect inspection apparatus to classify defects. Classification method.

25. An automatic defect classification method for inspecting an object to be inspected by a first defect inspection apparatus, and then inspecting the object to be inspected by a second defect inspection apparatus to classify defects occurring in the object to be inspected. ,

Inspection sample by the first defect classifier in the first defect inspection device Determining an inspection sample rate for each defect class according to the reliability of each defect class classified into

An inspection step of inspecting an inspection sample according to a sampling rate determined for each defect classification class for the inspection target in the inspection sample determination step for each defect classification class in a second defect inspection apparatus to classify defects. Automatic defect classification method.

26. An inspection information acquisition step of inspecting a defect sample group on an arbitrary sample by at least an arbitrary defect inspection apparatus to acquire sample inspection information, and a sample inspection information acquired by the inspection information acquisition step on a screen. Based on the display step for displaying the state of the defect attribute distribution of the defect sample group on the arbitrary sample based on the above-mentioned sample and the state of the displayed defect attribute distribution on the screen, the classification class element of the defect sample group is divided into branch elements. A decision rule setting step including a classification rule setting step of setting an individual classification rule for each branch element in a decision tree hierarchically arranged via a tree, and generating a defect classifier. Classifier generation process,

A defect classifying step of classifying the defect based on inspection information obtained by inspecting a defect generated in an object to be inspected by an arbitrary defect inspection apparatus using the defect classifier generated in the defect classifier generating step; A defect automatic classification method characterized by having a defect.

27. An inspection information obtaining step of obtaining a plurality of sample inspection information by inspecting a defect sample group on an arbitrary sample by a plurality of defect inspection apparatuses, and a plurality of sample inspection information obtained by the inspection information obtaining step. Are displayed almost simultaneously on the screen, and are reviewed to determine the classification class element of the defect sample group, and the classification class element is branched on the screen based on the classification class element determined in the review step. Hierarchically through elements A decision tree designating step of designating a decision tree to be expanded; and an individual decision element for each branch element in the designated decision tree based on at least one sample test information acquired by the test information acquiring step on a screen. A decision tree setting step including a classification rule setting step of setting a classification rule, and a defect classifier generating step of generating a defect classifier;

A defect classifier that classifies the defect based on inspection information obtained by inspecting a defect generated in an object to be inspected by at least one defect inspection apparatus using the defect classifier generated in the defect classifier generation process; Automatic defect classification method characterized by having a process.

28. At least one of the defect samples must be an image of at least one defect inspection device so that the same defect analysis sample for the same or the same type of defect sample becomes the same defect analysis reference. An automatic defect classification method, comprising: adjusting a processing procedure or image processing parameters; and classifying defects by a defect classifier based on the adjusted image.