JP2012173017A - Defect classification device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce network load and database load by quantitatively calculating significance of defects of images picked up and storing the images with defects having high significance alone in database, in a defect classification device for classifying types of the defects and foreign matters by picking up the images of the defects and foreign matters caused on substrates of integrated circuits, magnetic heads, magnetic discs, solar cells, optical modules, light emitting diodes, liquid crystal display panels and the like.SOLUTION: According to the present invention, defect coordinate data is input, an image is picked up in accordance with an image pickup program 501, a feature quantity of a defect is extracted from the image picked up in accordance with a feature quantity extracting program 502, the defect is classified by defect type in accordance with a defect classifying program 503, significance of each defect is calculated in accordance with a significance estimating program 504, and it is determined whether or not the image is to be transferred to database on the basis of the significance in accordance with an image selecting program 506.

Description

  The present invention calculates the importance of the defect from the feature quantity of the defective part extracted from the image data obtained by imaging the inspection object and the feature quantity of the pattern reflected as a background around the defective part. The present invention relates to a defect classification apparatus.

  In the manufacturing process of products manufactured using thin film processes such as exposure, development and etching, such as integrated circuits, thin film magnetic heads, magnetic disks, solar cells, liquid crystal displays, etc., in order to ensure high product yield, the manufacturing process It is necessary to detect and deal with various defects that occur at the early stage. This is usually done by the following steps: (1) A wafer to be inspected is inspected by a wafer appearance inspection device or a wafer foreign matter inspection device or the like, and the coordinates of the generated defect or the attached foreign matter within the wafer surface are detected. (2) Observe the detected defects and foreign matters (referred to as reviews), and classify the observed defects and foreign matters by type. Note that this review work usually uses a review-dedicated device having a microscope or the like for observing a defective part at a high magnification, but other devices having a review function, such as an appearance inspection device, may be used. . (3) For each classified result, collate it with the result of the electrical test, analyze the cause of the occurrence of defects and foreign matter, and take countermeasures.

  When the number of defects detected by the inspection device is large, the review work described above requires a great deal of labor. Therefore, an automatic review (Automatic Defect) that automatically captures defect images and automatically collects defect images. In recent years, development of review apparatuses having functions of “Review” and automatic defect classification that automatically classifies collected images has been active.

  Patent Document 1 and Patent Document 2 disclose a review apparatus and a manufacturing system having such an automatic review function and an automatic defect classification function. In Patent Document 1, a feature amount of a defective portion, for example, a feature amount such as a defect size or luminance is extracted from an image obtained by imaging the defective portion, and the image is used as a background of the defective portion as well as the defective portion. There is disclosed a technique for classifying defects by type by effectively utilizing feature amounts such as the shape and brightness of circuit patterns reflected together. In Patent Document 2, not only the defects are classified by type, but also the yield obtained as a result of the electrical test, that is, the ratio that can be shipped as a non-defective product among a plurality of integrated circuits formed on the wafer, and the defects are classified by type. A manufacturing system is disclosed that analyzes the correlation of results to determine the cause of each type of defect.

  Conventionally, the performance of the review apparatus, that is, the performance of the automatic review function and the automatic defect classification function is low, and it takes a lot of time to capture images of many defective parts. In such a situation, the technique disclosed in Patent Document 3 selects defects that seem to be highly important as pre-processing for the coordinates of defects detected by the wafer appearance inspection apparatus and the wafer foreign object inspection apparatus, and is important. The defect coordinates and the size of the defect output by the wafer appearance inspection device and the wafer foreign matter inspection device can be obtained without observing only defects that are likely to be high or by the technique disclosed in Patent Document 4 without performing a review. From the information, the results of electrical tests were predicted.

  In recent years, the performance of the review device, that is, the performance of the automatic review function and the automatic defect classification function has been improved, and it has become possible to capture images of a large number of defect parts in a short time. In such a situation, the technique disclosed in Patent Document 5 is used to quickly search for an image in which a defect of high importance that is likely to cause an electrical failure is reflected from a large number of captured images. It is effective to display instantly. In this technique, the degree of importance is defined by comparing the design layout data of the integrated circuit and the coordinates of the defect, and an image is searched using the size and importance of the defect. Patent Document 6 also deletes images that are not worth seeing in the future, such as images that do not have a correct defect due to some trouble with automatic review or images that are out of focus, from a large number of captured images. Thus, a technique for reducing the load on a database for managing captured images is disclosed.

JP 2001-331784 A US Pat. No. 6,408,219 Patent Publication 3678133 Japanese Patent No. 4357134 JP 2009-283854 A JP 2008-294361 A

The technologies described as background technologies can be broadly classified into the following three types.
(1) A technique for classifying defect types using various feature amounts obtained from an image.
(2) A technique for predicting the importance of a defect and limiting the image to be captured and a technique for predicting the yield before capturing an image with the automatic review function.
(3) A technique for displaying defects with high importance and a technique for deleting defects with low importance after taking an image with the automatic review function.

  One object of the present invention is to improve the performance (3). Specifically, according to the present invention, the feature amount of an image obtained from an image without the design layout data on the importance of the defect, the image captured by the automatic review function, without the design layout data. A defect classification device that quantitatively calculates importance using feature quantities such as pattern feature quantities and defect coordinates reflected as backgrounds.

  Further, the present invention provides a defect classification apparatus having a function of preferentially registering an image showing a defect having a high probability of causing an electrical failure in a database based on the importance of the defect calculated quantitatively. provide. With this function, it is possible to reduce the network load for transferring an image from the defect classification apparatus to the database and the load at the time of registration in the database. In addition, this function increases the probability that an image showing a defective part will exist for a product that is determined to be electrically defective when performing a correlation analysis between the result of the electrical test and the classification result of the defect. As a result, it is possible to speed up the cause analysis of the occurrence of defects.

In order to solve the above problems, for example, the configuration described in the claims is adopted.
The present application includes a plurality of means for solving the above-mentioned problems. For example, the application is connected to a database via a network, images of defects on the inspection object are taken, and the defects are classified into a plurality of types. A defect classification apparatus for predicting the importance of analyzing the defect, selecting an image of the defect having the high importance and registering it in the database, and a workpiece manufactured using a thin film process Means for inputting the coordinate information of the defect generated above from the inspection apparatus, image imaging means for imaging the defect on the workpiece according to the coordinate information, and the image for observing each input defect; On the other hand, the defect part and the background are separated by image processing, the feature quantity extracting means for extracting the feature quantity of the defective part and the background feature quantity, and using the extracted feature quantity, Classified into types Classifying means, defect result data obtained by observing defects detected from past inspection objects, and pass / fail data in chip units as a result of electrical testing of the inspection objects are associated with each other and grouped by defect category From the data, the importance prediction formula registration means for creating and registering a linear polynomial for each defect category with the feature quantity extracted from the defect observation image as an explanatory variable, with the ratio of non-defective product and defective product as an objective variable, Predicting means for predicting the importance of a defect by applying the feature quantity extracted by the feature quantity extracting means and the defect type classified by the classifying means to the importance predicting formula of the registered corresponding defect category; The defect classification apparatus includes a selection unit that selects whether to transmit the observation image to the database according to the importance.

  By applying the defect classification apparatus of the present invention in the manufacturing process of integrated circuits, thin film magnetic heads, solar cells, magnetic disks, liquid crystal displays, etc., it is possible to reduce the network load and database load in the factory.

It is a figure which shows an example of the outline | summary of the defect classification apparatus based on an electron microscope apparatus. It is a figure which shows an example of a series of flowcharts from an automatic review to image selection. It is a figure which shows an example of defect coordinate data. It is a figure which shows an example of the map which illustrated defect coordinate data visually. It is a figure which shows an example of a defect classification result. It is a figure which shows an example of the data which matched the defect classification result and the feature-value. It is a figure which shows an example of an importance degree prediction result. It is a figure which shows an example of the flowchart of importance prediction. It is a figure which shows an example of an importance prediction formula. It is a figure which shows an example of the flowchart which registers an importance prediction formula. It is a figure which shows an example of the result which matched the defect performance data read by step 601 and the electrical test data read by step 602 for every defect by step 603. FIG. It is a figure which shows an example of a selection condition setting screen. It is a figure which shows an example of the hardware constitutions of the inspection apparatus and database which are connected with the defect classification apparatus which concerns on this invention via a network. It is a figure which shows an example of the image figure showing the effect of this invention.

  First, as Embodiment 1, a defect classification apparatus based on an electron microscope for observing defects generated on a wafer will be described.

  FIG. 1 is a diagram showing an example of an outline of an electron microscope as an example of a defect classification apparatus according to the present invention. The electron microscope includes an electron source 201 that generates primary electrons 208, an accelerating electrode 202 for accelerating the primary electrons, a focusing lens 203 for converging the primary electrons, and a deflector 204 that deflects the primary electrons two-dimensionally. And an objective lens 205 for converging primary electrons on a substrate 206 such as a wafer. Reference numeral 207 denotes a drive stage on which the substrate 206 is mounted. 210 is a detector that detects the secondary electron signal 209 generated from the substrate 206, and 220a and 220b are backscattered electron detectors that detect the reflected electron signal 219, respectively. In the figure, two backscattered electron detectors 220a and 220b are installed facing each other, and each detects a different component of the backscattered electron signal emitted from the substrate 206. Reference numeral 211 denotes a digital conversion unit for digitizing the detected signal. Each of these parts is connected to the overall control unit 213 via the bus 218. In addition, the electron microscope includes a central processing unit (CPU) 214, a primary storage device 215, a secondary storage device 216, an input / output unit 217 such as a keyboard, a mouse, a display, a printer, and a network interface. Further, the secondary storage device 216 captures an image of a defective part generated on the substrate 206, that is, an image imaging program 501 for performing an automatic review, a feature amount of the defective part from the captured image, and a circuit pattern reflected in the background Feature amount extraction program 502 that extracts the feature amount of the image by image processing, defect classification program 503 that automatically classifies the defect by type using the extracted feature amount, and importance of the defect based on a pre-registered importance degree prediction formula Importance prediction program 504 for calculating degree, setting a prediction formula for importance, or setting conditions for selecting whether to register an image in a database via the network based on the calculated importance A selection condition setting program 505, an image selection program 506 for selecting an image based on the set conditions, and the importance prediction formula. Such importance prediction equation registration program 507 is stored to register. These programs are read from the secondary storage device 216 to the primary storage device 215 and calculated by the CPU 214.

  FIG. 2 is a diagram showing an example of a flowchart from input of defect coordinates to image selection according to the present invention including an automatic review function and an automatic defect classification function. In step 301, the coordinates of a defect or foreign matter (collectively, a defect) detected by the wafer appearance inspection device or the wafer foreign matter inspection device are input. In step 302, the variable N is initialized to zero. The variable N is a variable for repeating the loop process by the number of input defect coordinates. In step 303, the variable N is incremented. In step 304, if the variable N exceeds the number of defect coordinates read in step 301, the process flow is terminated. If not, the process proceeds to step 305. In step 305, an image of defect coordinates is taken. At this time, usually, an image is picked up with a low-power lens, the position of the defective part in the image is grasped by image processing, and then the high-power lens is switched, and the drive stage 207 is finely adjusted, and the defective part is To ensure that the image appears in the image. In addition, processing for focusing to capture an image is also performed. In step 306, feature amounts are extracted from the captured image. The feature amount separates the defective part and the background by image processing for the image, and quantifies the size, perimeter, aspect ratio, luminance of the defective part, and the luminance and shape of the circuit pattern in the background. It is a numerical value. Specifically, it is as described in Patent Document 1. In step 307, the defect is classified by type using the extracted feature quantity of the defective part, the feature quantity of the background, and possibly the coordinates of the defect in the integrated circuit. There are various definitions of defect types. For example, there is a method of dividing foreign matters, scratches, pattern defects, a method of dividing concave portions, convex portions, flat portions, and a method of dividing such as short circuit, disconnection, and isolation. In step 308, the importance of the defect is predicted. In the present invention, the importance is not the appearance of a defect that crosses between circuits. In the present invention, importance is defined as the probability of causing a defective product in an electrical test. That is, a regression equation for calculating the importance is registered in advance as a result of correlation analysis between the defect classification result and the electrical test result. In step 309, it is selected whether or not to transmit the image to the database via the network based on the calculated importance.

  FIG. 3 is an example of defect coordinate data input in step 301. The defect coordinate data 700 can be acquired as output data of a device that detects the position of a defect generated on the wafer, such as a wafer appearance inspection device or a wafer foreign matter inspection device. The defect coordinate data 700 is composed of a wafer number column 701 to be inspected, an X coordinate column 702 in the wafer coordinate system, and a Y coordinate column 703, in which defects for capturing images are arranged vertically. Here, an example is shown in which the coordinates of the defect are expressed in the wafer coordinate system, but the present invention is not limited to this. For example, a chip number is assigned to each integrated circuit formed on the wafer, and the coordinates of the defect may be expressed in a chip coordinate system in the integrated circuit. The present invention can define the defect coordinates by any method. I do not care.

  FIG. 4 is an example of the defect coordinate data 700 shown in FIG. 3 showing the position of the defect in the wafer coordinate system. A large round frame 401 represents a wafer, and an X axis 405 and a Y axis 406 in the wafer coordinate system are drawn in a state where a notch for defining the orientation of the wafer, that is, a notch 402 is on the lower side. Inside the wafer 401, a large number of integrated circuits 403 are drawn in a lattice pattern. In addition, a large number of black dots 407 represent the coordinates of each defect.

  FIG. 5 is a diagram illustrating an example of a result of classifying types for each defect in step 307. In the classification result 710, a defect type, that is, a defect category is added to the right end of each row of the defect coordinate data 700. Column 711 is the wafer number, column 712 is the X coordinate of the defect in the wafer coordinate system, column 713 is the Y coordinate of the defect, and column 714 is the defect category.

  FIG. 6 is an example of data in which defect coordinates, classification results, and feature quantities extracted in step 306 are associated with each defect. This is the data that is input to predict the importance in step 308. Corresponding data 720 is named as wafer number in column 721, X coordinate of defect in column 722, Y coordinate of defect in column 723, defect classification result in column 724, and F01 to F10 in column 731 to column 740, respectively. Features are shown. Here, ten types of features from F01 to F10 are shown as examples, but are not limited thereto. The feature amount is obtained by separating the defective part and the background by image processing on the image and quantifying the size, peripheral length, aspect ratio, luminance of the defective part, and the luminance and shape of the circuit pattern in the background. Although it is a numerical value, it is not restricted to this, For example, the numerical value obtained from other than an image like the coordinate of a defect may be utilized as a feature-value.

  FIG. 7 is an example of the result of predicting the importance in step 308. The predicted importance is added to the right end of the defect classification result 710 shown in FIG. The prediction result 750 shows the wafer number in the column 751, the X coordinate of the defect in the column 752, the Y coordinate of the defect in the column 753, the classification result of the defect in the column 754, and the calculated importance in the column 755. The importance is a numerical value of 0 or more and 1 or less because it is a probability of causing an electrical defective product. In addition to the image, the defect classification apparatus according to the present invention transmits data 750 to a database included in the yield management system, and stores the data 750 as useful information for analyzing the cause of failure. Of course, the feature amount of the defect shown in FIG. 6 is also transmitted to a database included in the yield management system and stored, which is useful information for analyzing the cause of the defect.

  FIG. 8 is a diagram showing an example of a flowchart of the importance level prediction step of Step 308, that is, the importance level prediction program 504. In step 521, the classification result and feature quantity data 720 shown in FIG. 6 are read. In step 522, the variable N is initialized to zero. The variable N is a variable for repeating the loop process by the number of defect coordinates of the input data. In step 523, the variable N is incremented. In the conditional branch of step 524, if the variable N exceeds the number of defect coordinates read in step 521, the process flow is terminated, and if not, the process proceeds to step 525. In step 525, a prediction formula is selected based on the defect category in column 724. In step 526, the values of the columns 731 to 740, that is, the feature amounts F01 to F10 are selected and substituted into the prediction formula, and the importance is calculated. In step 527, the importance is output as data 750 shown in FIG.

  FIG. 9 shows an example of a specific program for the step of selecting the importance prediction formula in step 525 and the step of calculating the importance in step 526. The program 530 first selects a prediction formula according to the defect category in the conditional branch of IF-THEN. If the defect category (variable defect_category) is “Budge”, the result of the linear polynomial having the feature values F01, F03, F06, and F07 as inputs is substituted into the variable x. If the defect category is “Dimple”, the result of the linear polynomial having the feature values F01, F03, F05, and F06 as inputs is substituted into the variable x. If the defect category is “Flat”, the result of the linear polynomial having the feature values F02, F03, F04, F06, and F08 as input is substituted into the variable x. Next, the value of the variable x is substituted into the sigmoid function, and the importance (variable Significance) is calculated.

  FIG. 10 shows an example of a flowchart of the importance prediction formula registration program 507 for registering the importance prediction formula used in step 525 and step 526. In step 601, defect record data is read. In step 602, electric test data for the wafer from which the defect record data has been collected is read. In step 603, defect record data including coordinate data for detecting the defect for each defect; Correlate with electrical test data of chips with defects. Next, in step 604, the number of defect categories included in the data associated in step 603 is extracted. For example, if three types of defect categories “Budge”, “Dimple”, and “Flat” are included, the number of categories is three. In step 605, a loop variable M for processing performed for each category is initialized. In step 606, the variable M is incremented. In the conditional branch of step 607, if prediction formulas are registered for all defect categories, the process proceeds to the end, and if there is an unregistered defect category, the process proceeds to step 608. In step 608, data for each defect category is extracted. In step 609, multiple logistic regression analysis is performed on the target defect category data. The multiple logistic regression analysis is a regression analysis for predicting the importance with a sigmoid function, as shown in the program 530 of FIG. 9, and has a feature of substituting the calculation result of a linear polynomial into a variable of the sigmoid function. In step 610, the importance degree prediction formula obtained as a result of the multiple logistic regression analysis is registered as in the program 530.

  FIG. 11 is an example of the result of associating the defect record data read in step 601 with the electrical test data read in step 602 for each defect in step 603. The associated data 760 includes a wafer number in column 761, columns 762 and 763, X and Y coordinates indicating the position of the defect in the wafer surface, column 764, a defect category, and columns 771 to 780, respectively. Defects obtained from the image or feature values of the pattern reflected as the background, column 781 is the electrical test data read in step 602 and associated with the chip unit, and the non-defective (Pass) or defective (Fail) Information. As can be seen from the wafer number in row 761, the data 760 was obtained for each defect detected from a plurality of past wafers, and was obtained in chip units as a result of an electrical test of these wafers. Information on good or defective products is associated. For example, when multiple logistic regression analysis is performed in step 609 for the defect category “Dimple”, in step 608, all the data in the row marked “Dimple” in column 764 is extracted, and in step 609, the column Multiple logistic regression analysis is performed using the ratio between the non-defective product (Pass) and the defective product (Fail) described in 781 as the objective variable and the feature values described in columns 771 to 780 as explanatory variables. However, it is not necessary to use all the feature values, and the feature values that are not significantly different between the non-defective product and the defective product are removed from the linear polynomial in advance, thereby improving the reliability of the prediction formula. In addition, defects and background feature values obtained from images may be used directly as feature values. However, principal component analysis is performed in advance, and the relationship between a plurality of feature values is orthogonal, that is, independent. You may keep it.

  FIG. 12 is an example of a graphical user interface of the selection condition setting program 505. The defect classification apparatus according to the present invention has a function of selecting an image in step 309 based on the importance output in step 527 in order to register an image obtained by capturing a defect in a database via a network. In the graphical user interface 800, an operation of a function for selecting an image is set. In the graphical user interface 800, the number of images to be transmitted to the database, the ratio, and the importance value can be set. When the check box 801 is checked and "50" is entered in the text box 804, 50 defects from the most important of the many defects that have been imaged by one automatic review function and automatic defect classification function The image is transmitted to the database. When the check box 802 is checked and "50" is entered in the text box 805, 50% of the large number of defects that have been imaged by one automatic review function and automatic defect classification function, from the one with the highest importance That is, if 200 defect images are taken, 100 defect images are transmitted to the database. Check boxes 801 and 802 cannot both be checked. In addition, when the check box 803 is checked and “0.3” is entered in the text box 806, the importance exceeds 0.3 among a large number of defects obtained by capturing an image with one automatic review function and automatic defect classification function. The image of the defect is transferred to the database. As shown in the example, if both the check boxes 801 and 803 are checked and “50” is entered in the text box 804 and “0.3” is entered in the text box 806, among the defects whose importance exceeds 0.3, The upper limit is 50 defects, and 50 images with the highest importance are transmitted to the database. Here, “transmit image to database” means that the defect classification device may send the image to the database by the push method, but the image is sent until the defect classification device requests the image transmission from the database by the pull method. It means to save in the specified directory. That is, image selection means storing images in a specified directory for transmission to the database.

  FIG. 13 is an example of a hardware configuration of an inspection apparatus or a database connected to the defect classification apparatus according to the present invention via a network. The defect classification device 810 is connected to a wafer appearance inspection device 811, a wafer foreign matter inspection device 812, an electrical test device (tester) 813, and a yield management system 814 via a network 815. The wafer appearance inspection device 811 and the wafer foreign matter inspection device 812 are sometimes referred to as a bright field inspection device and the wafer foreign matter inspection device 812 as a dark field inspection device because of differences in optical systems. The wafer appearance inspection apparatus 811 can scan a wafer using a high-magnification objective lens, process sequentially captured images, and detect a unique pattern. On the other hand, the wafer foreign object inspection apparatus 812 can illuminate from obliquely above the wafer and detect scattered light caused by unevenness on the wafer. Although the defects that are good at detection and the defects that are difficult to detect differ depending on the optical system, both can detect the position of the defect on the wafer. That is, both apparatuses are inspection apparatuses intended to output the defect coordinate data 700. The output defect coordinate data 700 is registered in the database of the yield management system 814 via the network 815. An electrical test apparatus (tester) 813 is an inspection apparatus that determines whether a plurality of integrated circuits on a wafer are electrically good or defective. Each integrated circuit is pre-fabricated with electrode pads for electrical testing. The electrical test apparatus 813 applies a prober included in the apparatus to an electrode pad of the integrated circuit on the wafer, flows various electrical waveforms for testing through the integrated circuit, and tests the performance of the integrated circuit. Test results, that is, good or defective data for each integrated circuit on the wafer are registered in the database of the yield management system 814 via the network 815. The yield management system 814 registers and manages various data sent from the wafer appearance inspection device 811, the wafer foreign matter inspection device 812, the defect classification device 810, and the electrical test device 813 in a database. The analysis tool of the yield management system 814 retrieves various data registered in the database and analyzes the cause of failure using statistical methods such as correlation analysis, regression analysis, analysis of variance, and hypothesis testing. it can. Further, the analysis tool can analyze the cause of the defect by drawing a wafer map as shown in FIG. 4 from the coordinates of the defect on the wafer and the position of the integrated circuit.

  FIG. 14 is a diagram showing an example of the effect when the defect classification apparatus according to the present invention is used as an example of the screen of the analysis tool of the yield management system. 820 is an example of a wafer map when the defect classification apparatus according to the present invention is not used, and 830 is an example of a wafer map when the defect classification apparatus according to the present invention is used. Large round frames 821 and 831 represent the wafer, and cuts for defining the orientation of the wafer, that is, wafer maps with notches 822 and 832 on the lower side, respectively are drawn. Inside the wafers 821 and 831, a large number of integrated circuits are drawn in a lattice shape. Among them, the squares 824 and 834 with thick frames are integrated circuits that have been determined to be defective by electrical tests, and squares 823 and 833 with thin frames. Is an integrated circuit determined to be a non-defective product by an electrical test. In addition, seven black dots 825, 826, 835, and 836 represent the coordinates of each defect whose image is stored in the database. Both of the wafer maps are the results of drawing the result of discriminating each integrated circuit into a non-defective product and a defective product in an electrical test, and displaying the image with the coordinates of the defect stored in the database superimposed thereon. This is because the automatic review and automatic defect classification functions picked up images of a large number of defects, but in order to reduce the network load and database load, only seven defect images were selected and stored in the database. It is an example. The difference between the two wafer maps 820 and 830 is that images of seven defects stored in the database are captured at different coordinates on the wafer. In other words, since the defects appearing in the stored image are different, the positions of the black dots 825, 826, 835, and 836 are different. When the defect classification apparatus according to the present invention is not used, an image stored in the database is randomly selected from a large number of captured images. Therefore, there is no correlation between the stored image and the non-defective product and the defective product of the electrical test. On the other hand, when the defect classification apparatus according to the present invention is used, the image stored in the database is selected and determined based on the importance predicted in the importance prediction program 504, that is, the importance prediction step 308. . Therefore, a correlation can be seen between the stored image and the non-defective product and the defective product of the electrical test. That is, there is a high probability that the black dots 835 and 836 overlap with an integrated circuit that is determined as a defective product by an electrical test, which is represented by a thick square. As a result, when analyzing the cause of the failure of the integrated circuit with the analysis tool of the yield management system 814, it is applicable when checking what kind of defect has occurred on the integrated circuit that has become defective in the electrical test. Images can be searched effectively.

  DESCRIPTION OF SYMBOLS 201 ... Electron source, 202 ... Accelerating electrode, 203 ... Condensing lens, 204 ... Deflector, 205 ... Objective lens, 206 ... Substrate (for example, wafer), 207 ... Drive stage, 208 ... Primary electron, 209 ... Secondary electron signal, 210 ... secondary electron detector 211 ... digital conversion unit 213 ... overall control unit 214 ... CPU 215 ... primary storage device 216 ... secondary storage device 217 ... input / output unit 218 ... bus 219 ... reflection Electron signal, 220a, 220b ... backscattered electron detector, 301 ... defect coordinate data reading step, 302 ... variable N initialization step, 303 ... variable N increment step, 304 ... conditional branching step, 305 ... image capturing program, 306 ... Feature amount extraction program, 307 ... Defect classification program, 308 ... Importance prediction program, 309 ... Image Separate program, 401 ... wafer, 402 ... notch, 403 ... integrated circuit, 405 ... X-axis, 406 ... Y-axis, 407 ... position of defect, 501 ... imaging program, 502 ... feature extraction program, 503 ... defect classification program 504 ... Importance prediction program, 505 ... Selection condition setting program, 506 ... Image selection program, 507 ... Importance prediction formula registration program, 521 ... Defect classification result and feature reading step, 522 ... Variable N initialization step 523 ... Variable N increment step, 524 ... Condition branch step, 525 ... Prediction formula selection step, 526 ... Importance prediction step, 527 ... Importance output step, 530 ... Prediction formula selection and importance prediction processing, 601 ... Defect record data reading step, 602 ... Electrical test data reading 603... Data association step, 604... Defect category number extraction step, 605... Variable M initialization step, 606... Variable M increment step, 607 ... Condition branch step, 608. 609 ... Multiple logistic regression analysis step, 610 ... Prediction formula registration step, 700 ... Defect coordinate data, 701 ... Wafer number, 702 ... X coordinate, 703 ... Y coordinate, 710 ... Defect classification result, 711 ... Wafer number, 712 ... X Coordinate, 713 ... Y coordinate, 714 ... Defect category, 720 ... Feature amount extraction result, 721 ... Wafer number, 722 ... X coordinate, 723 ... Y coordinate, 724 ... Defect category, 731 ... Feature amount F01, 732 ... Feature amount F02 , 740 ... Feature amount F10, 750 ... Importance prediction result, 751 ... Wafer number 752: X coordinate, 753 ... Y coordinate, 754 ... Defect category, 755 ... Importance, 760 ... Corresponding data of defect result data and electrical test result, 761 ... Wafer number, 762 ... X coordinate, 763 ... Y coordinate, 764 ... defect category, 771 ... feature quantity F01, 780 ... feature quantity F10, 781 ... electrical test result, 800 ... selection condition setting graphical user interface, 801-803 ... check box, 804-806 ... text box, 810 ... defect classification device 811: Wafer appearance inspection device, 812 ... Wafer foreign matter inspection device, 813 ... Electrical test device (tester), 814 ... Yield management system, 815 ... Network, 820 ... Wafer map, 821 ... Wafer, 822 ... Notch, 823 ... Non-defective product Integrated circuit, 824... Defective integrated circuit, 825, 826. The position of the image there defects, 830 ... wafer map, 831 ... wafer, 832 ... notch, 833 ... non-defective integrated circuit, 834 ... defective integrated circuits, 835 and 836 ... position of the image there defect

Claims (8)

Connected to a database via a network, captures images of defects on the inspection object, classifies the defects into a plurality of types, predicts the importance of analyzing the defects, A defect classification device for selecting images and registering them in the database,
Means for inputting coordinate information of defects generated on a workpiece manufactured using a thin film process from an inspection apparatus;
Image input means for imaging the defect on the workpiece according to the coordinate information, the input image for observation of each defect;
For each of the images, a defect part and a background are separated by image processing, a feature quantity extracting unit that extracts a feature quantity of the defective part and a background feature quantity, and
Classification means for classifying the defect into a plurality of types using the extracted feature amount;
Corresponding defect result data obtained by observing defects detected in the past inspected object and good / bad data in chip units as a result of electrical testing of the inspected object. Importance prediction formula registration means for creating and registering a linear polynomial for each defect category with the feature quantity extracted from the defect observation image as an explanatory variable, with the ratio of defective products as an objective variable,
Prediction means for predicting the importance of a defect by applying the feature quantity extracted by the feature quantity extraction means and the defect type classified by the classification means to the importance prediction formula of the corresponding defect category registered. When,
A defect classification apparatus comprising: sorting means for sorting whether to transmit the observation image to a database according to the importance. A defect classification device that takes an image of a defect on an inspection object, classifies the defect into a plurality of types, predicts the importance of analyzing the defect, and outputs the importance corresponding to each defect. There,
Means for inputting coordinate information of defects generated on a workpiece manufactured using a thin film process from an inspection apparatus;
Image input means for imaging the defect on the workpiece according to the coordinate information, the input image for observation of each defect;
For each of the images, a defect part and a background are separated by image processing, a feature quantity extracting unit that extracts a feature quantity of the defective part and a background feature quantity, and
Classification means for classifying the defect into a plurality of types using the extracted feature amount;
Corresponding defect result data obtained by observing defects detected in the past inspected object and good / bad data in chip units as a result of electrical testing of the inspected object. Importance prediction formula registration means for creating and registering a linear polynomial for each defect category with the feature quantity extracted from the defect observation image as an explanatory variable, with the ratio of defective products as an objective variable,
Prediction means for predicting the importance of a defect by applying the feature quantity extracted by the feature quantity extraction means and the defect type classified by the classification means to the importance prediction formula of the corresponding defect category registered. When,
A defect classification apparatus comprising: output means for outputting the predicted importance in association with each defect. The importance level prediction formula registration means
Corresponding defect result data obtained by observing defects detected in the past inspected object and good / bad data in chip units as a result of electrical testing of the inspected object. 3. A multiple logistic regression equation for each defect category is created and registered with the ratio of defective products as an objective variable and the feature quantity extracted from the observed image of the defect as an explanatory variable. The defect classification apparatus described in 1. The feature amount extraction means is
2. A defect part and a background are separated from each image by image processing, and a defect position is extracted as a feature quantity in addition to a defect part feature quantity and a background feature quantity. Or the defect classification device according to claim 2. The prediction means is
The feature amount extracted by the feature amount extraction unit and the defect type classified by the classification unit are applied to the importance prediction formula for the registered corresponding defect category and calculated, and the calculation result is a variable of the sigmoid function. The defect classification apparatus according to claim 1, wherein the importance of a defect is predicted by substituting into.   The claim according to any one of claims 1 to 5, wherein the feature amount is a feature amount obtained by converting a plurality of feature amounts extracted from an image into an orthogonal state by principal component analysis. The defect classification device according to item.   The defect classification apparatus according to claim 1, wherein the selection unit receives an upper limit setting for the number of defects transmitted to the database. Step of inputting defect coordinates data obtained by observing defects detected from the past inspection object, defect coordinates, feature amount extracted from image, defect type data, and
Determining the inspected object with an electrical test device, and inputting electrical test data having information on a good product or a defective product;
Associating the defect record data with the electrical test data for each chip;
Extracting data associated with each defect type;
And a step of executing a logistic regression analysis on the extracted data using the electrical test data as an objective variable and the feature quantity as an explanatory variable.

Images