JP2004295879A - Defect classification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a defect classification method capable of estimating propriety, when defects are automatically classified and setting the classification standards which assure classification performance improvement, with less labor. <P>SOLUTION: Defects of unknown classification class are classified based on two different classification standards, defects varying in a classification result are collected, and defect classification classes are manually given to the collected defects. The defects, given the classification classes, are divided into a classification standard setting group and an evaluating group; classification performance is calculated when the defects belonging to the evaluating group are classified based on the classification standard set so as to classify the defects belonging to the classification standard setting group with maximum performance; and the propriety of the set classification standard is evaluated. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

The present invention relates to a defect classification method for industrial products, and more particularly to a method for setting an automatic classification standard in an automatic defect classification method after a pre-process inspection of a semiconductor product, in which it is important to perform detailed review and classification of the defect after detecting the defect.

  With the miniaturization of semiconductors, it is becoming increasingly difficult to control the pre-process of the semiconductor manufacturing process, and in the conventional process management based on the variation in the number of semiconductor defects detected by visual inspection of semiconductor wafers, it is high. Semiconductors cannot be manufactured with a high yield. Therefore, after the inspection by the visual inspection device, an image obtained at the time of the inspection is analyzed to automatically classify the defect, or a more detailed image of the defective portion is re-detected after the visual inspection, and the image is automatically classified. ADCs have been proposed.

  The ADC has a rule type that classifies defect features composed of multiple image features, such as brightness extracted from images and defect shapes, into defect classes based on predetermined rules. Are defined as one scalar value, a plurality of the scalar values are collectively converted into a multi-dimensional vector, and a criterion for classifying defects based on a defect class distribution in a multi-dimensional space spanned by the multi-dimensional vector is automatically determined. Various methods have been proposed, such as a teaching type to be generated, and a combination of a rule type and a teaching type.

  In order to automatically classify defects by the ADC, it is necessary to set defect classification criteria based on a defect sample whose classification class is known before performing automatic classification. In the case of the rule type, it is generally necessary to set judgment thresholds for some items of the defect feature. In the case of the teaching type, it is required to obtain the distribution of defect classes in the multidimensional space.

  As related to the above-mentioned prior art, as a method of setting a classification criterion, a rule-type method, a method of teaching a feature amount, and a method of combining these are known. For example, as a classification method of a rule format, there is a method described in Patent Document 1, and as a method of teaching a feature value, Patent Document 2 describes the classification method and the teaching method. Further, a classification method combining the rule type and the teaching type combined with each other is described in Patent Document 3.

JP 2001-331784 A JP 2001-256480 A JP 2001-135682 A

  In the prior art, since the defect classification criteria are set based on the distribution of the defect features before the automatic classification is started, the distribution of the defect characteristics after the automatic classification operation is no longer similar to the distribution before that. In such a case, there is a problem that appropriate automatic classification cannot be realized. For example, if some important defect samples did not occur in the defect samples prepared when the classification criteria were set, the classification criteria of the defect class for which the samples could not be prepared would not be set. Therefore, automatic classification cannot be performed normally as it is. When starting automatic classification, it is unknown whether defects of a different type from the defect samples already collected occur in the target semiconductor layer, so the classification criteria set at this time are sufficient It is difficult to determine whether.

In order to solve this problem, finding the defect class sample that was insufficient when setting the classification criteria from the defects processed in large quantities after the start of the automatic classification, and updating the classification criteria is the best classification performance. Needed to achieve. However, constantly monitoring the state of defect classification after the start of the automatic classification requires a lot of effort, so that the meaning of the automatic classification is lost, and this operation cannot be performed in an actual production line. As a conventional invention for solving this problem, as disclosed in Patent Document 1, for example, a feature amount is calculated from a defect detected on a semiconductor wafer and assigned to a feature amount space. , Assigning categories from the distribution of the data, creating teaching data for defect classification, and performing defect classification, performing similar processing on the defects to be classified detected from the semiconductor wafer and comparing the data with the teaching data. When a difference occurs in these data due to a variation in the process, a method has been proposed in which the previously obtained teaching data is corrected according to the difference.
However, recent process management requires fine classification of defects, and the number of dimensions in the feature space has been increased to achieve this. It is more difficult to automatically detect a change in the feature amount distribution. As a result, the present invention has not been put to practical use until now. Further, even if the defect state after the start of the automatic classification is monitored and it is found that there is a problem in the setting of the classification criterion, it is actually extremely difficult to reset the classification criterion.

  In general, it is not sufficient to guarantee that a large amount of classification results that have been processed after the initial setting of classification criteria will result in more accurate classification results by resetting the classification criteria. This is because the true classification class of the automatically classified defect is impossible in an unknown state. Also, before and after resetting the classification criteria, the classification results may fluctuate differently, and before and after deflection, the number of defects for each classification class may fluctuate. In addition, there is a problem that inconvenience occurs in performing the above.

  Further, there is a problem that a general user does not know how to reset the classification criteria to improve the overall classification performance. In resetting the classification criterion, it is general to calculate a feature amount of a defect belonging to each defect class in a prepared defect group and set a criterion based on this distribution. If a feature amount is extracted for a defect, the performance may be deteriorated particularly when some defects are added. Therefore, when trying to obtain the best performance, it is important to select only the defects necessary to obtain the best classification result before changing the classification criterion. It is difficult for a general user to judge whether or not a result can be obtained. After all, the distribution of the feature amount of the defect is calculated from all the defects, and the classification criterion is set. There was a difficulty that improvement could not be realized at all.

  SUMMARY OF THE INVENTION An object of the present invention is to provide a defect classification method capable of estimating the validity at the time of automatically classifying defects with a small amount of labor and setting a classification standard that guarantees an improvement in classification performance.

  Another object of the present invention is to realize a method of resetting a classification criterion that can guarantee an improvement in overall classification performance.

In order to achieve the above object, according to the present invention, a method of classifying defects includes a step of classifying an image of a first defect group detected by imaging a sample for each classification class of defects, Creating a classification criterion using information on the classified images of the first defect group;
Classifying an image of a second defect group on the sample having a known classification class based on the created classification criterion, and classifying the image of the second defect group having a known classification class based on the classification criterion The method includes a step of obtaining a classification accuracy rate of the classification criterion based on a result of the classification, and a step of correcting the classification criterion based on the data of the classification accuracy rate.

  Further, in order to achieve the above object, according to the present invention, a method of classifying defects includes a step of classifying a group of defects detected on a sample for each defect classification class to generate first teacher data; Creating a classification criterion using the information of the first teacher data; categorizing a part of the first teacher data based on the classification criterion; Obtaining the classification performance of the classification criterion; and removing second defect data from the first teacher data from the first teacher data to generate second teacher data. Classifying defects whose classification class is unknown using the second teacher data.

  Further, in order to achieve the above object, according to the present invention, a classification criterion of a defect is created based on teacher data with a known class, and the classification class of the defect is classified based on the classification criterion with the known data. A method for classifying a defect that corrects the classification criterion based on the determined classification performance based on the classification performance of the defect based on a result of classifying the known data based on the classification criterion based on the classification criterion, In the step of creating the classification criterion, the classification criterion is set using a classifier having a multilayer structure including an upper layer, an intermediate layer, and a lower layer, and the classifier of the intermediate layer is determined to be classified by a logical expression. The logical formula is set so that the classification correct rate by the lower classifier is minimized when the logical formula is changed.

  ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to estimate the validity at the time of automatically classifying a defect, and to set the classification criterion which guaranteed the improvement of the classification performance with little effort. Further, it is possible to realize a method of resetting the classification criterion that can guarantee the improvement of the overall classification performance.

  Hereinafter, an embodiment of the present invention will be described with reference to FIGS.

A semiconductor wafer is patterned in a multilayer structure by a number of steps. In the process of manufacturing the multilayer structure, in order to monitor the manufacturing process, a visual inspection for each layer, a review of defects detected in the visual inspection, and a classification for each defect type are performed. Defect review methods generally include:
(1) An inspection device reviews images taken during inspection,
(2) A revisit-type one that captures an image of a defect again using the external appearance image pickup device of the inspection device;
(3) Several methods have been proposed and used, such as a method using a review device such as a review SEM (scanning electron microscope), which is different from an inspection device.

  Here, the sequence of the type (3) will be described in more detail. After the appearance inspection is performed, the appearance of the defect detected in the inspection target at the coordinates on the wafer is again imaged by the review device to obtain the defect image. Then, the user manually classifies the image into defect classes such as foreign matter, pattern defects, and scratches, and calculates the number of defects, the distribution of defect sizes, and the distribution of defect occurrence positions on the wafer for each defect type. Analyzes and finds problems in the manufacturing process.

  As recent semiconductor processes become more and more miniaturized, the manufacturing process becomes more frequent due to a slight deviation from the optimal state, and a larger number of defect images need to be reviewed. ing. Therefore, recently, ADCs that automatically classify the reviewed defects instead of classifying the defects manually are increasing in importance.

  FIG. 1 shows a classification sequence of the ADC according to the present invention. Step 101 is a manual classification operation step, in which a plurality of collected defect images are visually evaluated and classified. Step 102 is a step of setting a classification criterion using the defects visually evaluated in step 101. The internal sequence of step 102 will be described in detail with reference to FIG. 2. If it is determined in step 101 that the number of defects manually classified in step 101 is small, defect images are collected again and step 101 is executed.

  Step 103 is an image acquisition step in which a defect image to be classified is acquired using the classification criterion set in step 102. As an acquisition method, a method of transferring already captured image data to a memory area for executing a program by means such as communication, or a method of newly capturing an image by a detection system as shown in FIG. 7 described below. Is applicable. Step 104 is an automatic classification step for classifying a defect whose classification result is unknown.

  Step 105 is a check necessity determination step, in which it is determined whether it is necessary to evaluate the classification criteria set in step 102. One embodiment of step 105 will be described in detail with reference to FIG. If it is determined in step 105 that the evaluation of the classification criteria is unnecessary, it is determined in step 108 whether or not the classification sequence has been completed. If it is determined that the processing has not been completed, the process returns to step 103.

On the other hand, if it is determined in step 105 that the classification criterion needs to be evaluated, an automatic classification result check step in step 106 is executed. Step 106 is a step of manually evaluating a part of the defects automatically classified in step 104 in order to evaluate whether the automatic classification has been correctly performed in step 104. Here, if it is determined that the automatic classification has been correctly performed, the step 103 for capturing an image whose classification result is unknown and the step 104 for automatically classifying the image are further performed. As a result of the check in step 106, if the accuracy rate of the automatic classification result is low, the classification criterion is set again in step 107 using the image manually evaluated in step 105, and the process returns to step 103 to return to step 103. Execute the sequence up to the automatic classification result check Before describing this sequence in more detail, a method for automatically classifying defects will be described. In order to classify a defect, it is necessary to first extract the characteristics of the defect. Defect characteristics include image characteristics and material composition characteristics as typical ones. The image feature is obtained by comparing two images of an image having a defect and an image of an area having the same pattern without a defect, calculating a defect area, and then calculating a defect size and the like from the calculated defect area. Methods for calculating shape features, lightness of defects, image texture, and surface irregularities have been proposed. The material composition feature includes, for example, a spectrum obtained from an energy dispersive X-ray fluorescence spectrometer (EDX) that irradiates a defect with an electron beam and analyzes the X-ray.

  As a method of setting a classification criterion, a rule-type method, a method of teaching a feature amount, and a method of combining these are known. For example, as a classification method of a rule format, there is a method described in Patent Document 1, and as a method of teaching a feature value, Patent Document 2 describes the classification method and the teaching method. Further, a classification method combining the rule type and the teaching type combined with each other is described in Patent Document 3.

  For example, in order to classify concave and convex rules and others in the rule type, the method disclosed in Patent Document 1 discloses a method in which a shading direction obtained from left and right images and a difference between a defect image and a reference image are different. It is described that it is possible to determine whether a defect is convex, concave, or neither by using position information of a defective portion obtained from an image. In general, since minute noise is generated in an image, a threshold value for determining whether the image is a shadow or a noise caused by a flaw is required. For example, a criterion for classifying in such a manner that a feature such as a degree of unevenness obtained by summing up the intensity of the shadow over the entire defect area is calculated, and if this exceeds a certain threshold value, there is unevenness, otherwise the surface is flat is set. Is done. In order to set the threshold value indicating the classification boundary, the attributes of the defects, that is, the defects whose attributes such as flat, convex, and dent are known in advance in this case, are collected in advance, and these are collected. It is common to make the most appropriate classification.

  As a classification using the teaching type, for example, as shown in Patent Document 2, there is a classification that forms a feature amount space for teaching based on feature amounts extracted from a defect image. Here, c is a feature amount extracted from one defect image, n is a predetermined number of feature amounts, and a combination (c0, c1,..., N) of n feature amounts extracted from one image. cn) is used as a feature vector, and a point in the feature space represented by the feature vector extracted from the defect image used for calculating the feature space is used as a teacher point, and a discriminant for determining these teacher points is calculated. How to do is described. As described above, in order to form a feature amount space, it is necessary to collect a plurality of defects whose classification classes are known and set a teacher point.

  As shown in these known examples, in order to perform the step of setting the classification criterion shown in step 102, it is generally performed to collect defects in advance and to set them from the group of defects. Conventionally, when a certain number of defects is collected, a classification standard is set based on the collected number. However, depending on the layer of interest, it may be necessary to collect a large number of defects in order to perform automatic classification correctly, while ensuring accurate classification regardless of how many images are collected. Some layers do not allow you to set a good classification criterion. For this reason, if the classification performance is not sufficient even after setting the classification criterion once, it is difficult to determine whether the layer cannot be classified automatically or the number of acquired images is insufficient. In this case, it was necessary to collect images and reset the classification criteria again.

  FIG. 2 shows a classification reference setting sequence according to the present invention. First, in step 201, a value N for dividing each defect group collected in step 101 into a certain number N of groups is set. As the initial value of N, 2 is convenient. Next, in step 202, the defects of each defect group are divided so as to be included in the N groups. The true classification class of each defect is manually given in step 101, and it is preferable to divide each defect group so that each defect type is as equal as possible. Next, in step 203, one of the N groups is selected as a classification group, and the remaining N-1 groups are selected as teaching groups, and classification criteria are set based on the defects belonging to the teaching group.

  Next, in step 204, defects belonging to the classification group are classified according to the set classification criteria. Since there are N combinations of selecting one group as a classification group from N groups, the steps from step 203 to step 204 are repeated until all the combinations are completed. When all of the N combinations are completed, a classification result of all the defects is obtained based on the classification criterion obtained from the teaching set to which the defect to be classified does not belong. Next, in step 201, the value of N is increased and the process is executed again. N may be implemented up to the number of defects when the number is small, or up to about 10 to 20 when the number is large.

  Now, the number of defects used for setting the classification criterion in the defect class i is set to Ci, and is referred to as a teacher defect. When N is performed from 2 to 20, the correct classification rate of the automatic classification when the number of teacher defects changes from Ci / 2 to Ci * 19/20 can be obtained. In general, as the number of teacher defects increases, the accuracy rate for classifying the defects that are not used for setting the classification criterion increases, so that the relationship between the teacher error number and the accuracy rate as shown in FIG. 3 is obtained. Further, from the classification accuracy rate when the number of teacher defects is set to the maximum N, it is possible to estimate the classification performance when all the defects given in step 101 are taught. If the classification correct answer rate is sufficiently high, it is considered that the correct answer rate is high when any defect whose classification result is unknown is automatically classified based on the calculated classification criterion.

  On the other hand, when the correct answer rate is low, two cases can be considered. One is a case where the number of teacher defects is insufficient, and the other is a case where classification with high performance is originally impossible with the ADC. This difference can be determined by the curve shown in FIG. As shown in FIG. 3, if the differential value of the curve at the time when N is maximized takes a positive value, the performance can be expected to be improved by increasing the number of teacher defects. Can be determined. On the other hand, if the differential value of the curve at the time when N is maximized is near 0, the classification accuracy rate is already saturated, and even if the number of teacher defects is increased, the performance cannot be improved. It turns out that ADC cannot expect high performance.

  In step 205, a graph of each class class and the number of teacher defects of the total performance-classification accuracy rate is displayed, and if the performance is not sufficient, it is necessary to increase the number of teacher defects, or Display a report on what you can't expect. If the performance is not sufficient based on the displayed report, the number of teacher defects is increased, steps 201 to 204 are repeated, and the evaluation result is displayed again in step 205. After the classification accuracy rate evaluation is completed in this way, in step 206, a classification criterion is generated using all teacher deficiencies, and this is used as the classification criterion generated in step 102.

  Next, step 104 will be described. Although the teacher defects collected when performing the step 102 are guaranteed to be classified with high performance according to the sequence shown in FIG. 2, only the same defects are generated for a long time thereafter. Therefore, there is a problem that the performance cannot always be maintained continuously. Therefore, it is necessary to monitor whether the defect is normally classified by manual classification every certain period. As this method, it is desirable to incorporate a monitoring rule into a classification standard for performing ADC.

Monitoring rules are based on the following external inputs:
(1) Time-series variation in the number of defects existing per wafer in the wafer group where the defect group was detected. (2) Defect position space obtained from the defect group detected for each wafer in the wafer group where the defect group was detected. (3) Number of teacher defects used in setting the classification criteria (4) Electrical test results applied to the wafer group where the defect group was detected (5) For some or all of the defect group The monitoring timing or defect group is determined based on at least one result of the defect composition analysis performed.

  Generally, it is considered that a defect that occurs stably is already present in the image collected in step 101 when setting the classification criterion, and it is necessary to pay particular attention to a defect that occurs suddenly here. . From this point of view, as described in (1), the time-series variation of the number of defects existing per wafer is analyzed, and when the number of defects existing per wafer becomes extremely large, the classification result should be monitored. For example, it is possible to evaluate a classification result of a defect that rarely occurs that does not occur when the process is in a stable state, and it is easy to incorporate the result into a teacher defect group. In the spatial distribution of defect positions described in (2), it is known that random defects, systematic defects, and the like can be obtained from the distributions. In such a case, by monitoring the classification result, it is possible to confirm the classification state of a defect having a characteristic different from that of the taught defect.

  If a large number of teacher defects already exist, there is little need to monitor the classification state. If the number of teacher defects is small, it is necessary to check the classification state every short period.

  One of the main purposes of defect classification is to identify the cause that is reducing product yield. As a method of identifying the cause of a defect, an electrical test and a composition analysis of the defect are often performed on an actual production line. Here, it is necessary to classify a defect having high electrical or physical fatality with high accuracy, and to estimate a defect cause and a yield with high accuracy by confirming these classification results. A possible automatic classification can be realized.

  This is illustrated in FIG. Reference numeral 801 denotes a classification reference confirmation timing determined in advance when setting the classification standard. If the number of defects used in setting the classification criterion in step 102 is large, the confirmation timing is set long, and conversely, if the number is small, the confirmation timing is set short. Further, when the monitored results of the above (1) to (5) greatly fluctuate, the confirmation timing is changed according to the fluctuation. Reference numeral 802 denotes a time-series variation in the number of defects detected per wafer described in (1). Reference numerals 803 and 804 denote the number of defects detected at a position where a characteristic pattern is generated in the distribution of defect positions detected on one wafer, 803 denotes the number of defects at a position indicating the distribution of defect positions of a linear pattern, Reference numeral 804 denotes a distribution of the number of defects having an annular distribution. Reference numeral 805 denotes a change with time in the number of chips determined as defects as a result of the electrical test, and reference numeral 806 denotes the number of defects in which iron was present among the elements detected by EDX analysis. Here, before the classification result confirmation timing determined from the number of teaching defects indicated by reference numeral 801, it was confirmed that a large number of annular defects that had not occurred before that were generated. The ADC system informs you that the classification results should be confirmed. At this time, the ADC system automatically extracts a defect group for which the defect classification result should be confirmed from the defects classified in the period up to 807.

  In this manner, the ADC system extracts a defect group whose classification state is to be confirmed from the defects automatically classified in step 104 according to the monitoring rule set in advance, and displays the defect group to the user. This processing may be performed by, for example, a review device that captures a defect image, or may be performed by a computer terminal connected to a network. The user gives a manual classification result to the defect group extracted by the ADC system, whereby the automatic classification result check step shown in step 106 is realized. The classification result checked in step 106 is desirably used as it is for resetting the classification standard.

  However, when the classification criterion is changed here, a problem arises in that the use of the teaching data obtained in step 106 cannot guarantee that the classification performance is truly improved. That is, the teacher defect used at the time of initial setting and the defect group confirmed in step 106 are both used as teacher defects, and a new classification criterion is generated in step 107. Therefore, the classification criterion according to the sequence described with reference to FIG. Can be confirmed in the teacher defect group when using, but the true classification result is unknown for defects other than those already classified in step 104 and the classification results are manually given in step 106, and It is not possible to confirm whether the performance will be improved or degraded according to the classification criteria set in. In fact, many phenomena have been reported in the classification engine that the classifying performance is degraded by giving a biased teaching. In order to generate a classification criterion in step 107, it is necessary to guarantee that the classification performance is not degraded. Is done.

  A solution according to the invention for this is shown in FIG. FIG. 4 shows the processing performed in step 107 of FIG. Step 401 is a step of setting a classification criterion using the given teacher defect. Step 402 is a step of reducing all or a part of the defects classified in step 104 by both the pre-change and post-change classification criteria. This is a step of classifying and extracting a difference between the classification results. By confirming only those extracted as having a difference in the classification result in step 402, it is possible to perform performance evaluation before and after the change of the classification standard for all the defects compared in step 402.

  In step 403, the classification is performed in step 402, and the true classification result is manually taught only for those having a difference in the classification result. If the initial classification criterion generated in step 102 has already been sufficiently adjusted, the number of cases in which there is a difference in the classification result is usually small. The true classification result input in step 403 and the result classified in step 402 are compared and evaluated. If the classification performance based on the classification criterion generated in step 107 is lower than the initial classification criterion generated in step 102, In step 403, the manually classified defect is added to the teacher defect group, and in step 401, the classification criterion is set again. By forming such a sequence, it is possible to set a classification criterion that is guaranteed to improve performance by adding a teacher defect.

  So far, only the method of adding a teacher defect to improve the classification performance has been described. However, in order to further improve the classification performance, it is necessary to remove unnecessary teacher defects. The teacher defect image is normally captured normally, but in some images, the defective area itself cannot be properly extracted depending on the image capturing state such as charging of the SEM image, or abnormal brightness due to charging. In some cases, it is detected. When the classification criterion is set using the defects imaged in this way, the criterion is reset using the feature amounts that should not be originally distributed, so that the classification performance is rather deteriorated.

  Therefore, it is necessary to extract teacher data that lowers the classification performance. This one execution sequence will be described with reference to FIG. First, in step 501, one defect for which necessity is determined is designated from the teacher defect group. Next, in step 502, the teacher defect except for the defect designated in step 501 is divided into N groups. Here, N is equal to or less than the number of teacher defects, but may be divided into, for example, about 20. Here, each group is divided such that the number of defects belonging to each defect class is as equal as possible.

  In step 503, one of the groups is selected, a classification criterion is calculated from the defects belonging to the non-selected group, and in step 504, the defects belonging to the group selected in step 503 are classified. If step 503 and step 504 are performed N times while changing the group to be selected, a classification result of all teacher defects can be obtained. In step 505, defects other than teacher defects for which true classification results are known are classified. The classification results obtained in steps 504 and 505 are compared with the true classification results, and in step 506, the classification performance after removing the teacher defect specified in step 501 is compared with the classification performance before the removal. You. In step 506, when it is determined that the performance is improved by removing the defect from the teacher defect, the defect is a candidate that is better to be deleted than the teacher defect.

  In addition, as a method of confirming the improvement of the classification performance, there are a method of comparing the accuracy rate of each classification class, a method of comparing the total performance, and a method of combining these. In both comparisons, if the performance is improved, it is a defect that must be deleted, if only one of the performances is degraded, it can not be said either, and if both are degraded, it is an essential defect as a teaching defect. become.

  By performing the sequence of steps 501 to 506 while changing the teacher defect selected in step 501, unnecessary teacher data is extracted.

  Next, in step 507, the rankings are ranked in the order of remarkable improvement in the performance in the comparison in step 506, and in step 508, the teacher defect groups are deleted one by one from the top, and in step 509, the remaining teacher defect groups are deleted. Classification criteria are set.

  Step 510 is a step of performing classification based on the classification criteria set in step 509, and internally performs the processing described in steps 502 to 505.

  Next, in step 511, the classified result is compared with the true classification, and if an improvement in performance is recognized, the processing in step 508 is performed again. On the other hand, if the performance does not improve, the teacher defect which has been finally removed from the teacher defect group in step 512 is returned to the teacher defect, and the process returns to step 508 to evaluate the next defect.

  The sequence described with reference to FIG. 5 is an extremely important process in determining whether to add a small number of teacher defects. This is because, when a small number of teacher defects are added, the feature amount distribution of the teacher defect group to be added does not necessarily match the feature amount distribution of the defect class. Therefore, incorporating the sequence shown in FIG. 5 in step 401 or step 403 of the sequence shown in FIG. 4 is important for setting the best classification criterion.

  Here, the sequence shown in FIG. 5 is a method that guarantees that the overall performance is improved, but on the other hand, even if the overall performance is slightly reduced for the user, it is extremely important to classify it. Defects exist and care cannot be taken to improve the classification performance of this special case. For this reason, in order to improve the performance by the sequence shown in FIG. 5, it is displayed that the user can clearly see the defect that is desirably deleted from the teacher defect group so that the user can confirm the defect. It is desirable to be able to compare the performance when deleting each defect and when deleting each defect.

  Now, with respect to a method of setting a classification criterion, in the case of a teaching-type classifier, Patent Document 2 describes a method of setting the classification criterion. On the other hand, as described in Patent Document 3, a classifier of the first layer and a classifier of the lower layer, such as a classifier in which knowledge of a semiconductor is shown in advance by rule logic, are configured. In such a case, it has been conventionally required to manually set parameters for setting the classification criteria of the first layer. Even if the classifier of the lower layer is taught, the classifier of the lower layer and the classifier of the first layer do not necessarily correspond to 1: 1. This is because a classification result cannot be obtained.

  In order to solve this problem, the method according to the present invention solves this problem by manually giving a classification result even to an intermediate classification result for an upper classifier, that is, a user who sets classification criteria. Corresponding. FIG. 6 shows an example of this. Reference numeral 601 denotes a GUI (Graphical User Interface) for setting a classification criterion, and reference numeral 602 denotes a tag indicating a class of a defect in a lower layer. A tag 603 indicates a defect class of the intermediate layer. In order to classify an arbitrary defect into the defect class CB (i) in the lower layer, provided that the defect is classified into the class CM (j) in the intermediate layer, CB (i) is displayed on the GUI. Are linked with the tag 602 corresponding to CM (j).

  In the area 604, a thumbnail image 605 of the defect is displayed. The GUI 601 allows a user to input and instruct with a mouse pointer. When the user designates the tag 602 with a pointer, a thumbnail image 605 of a defect belonging to the defect class indicated by the tag 602 is displayed in the area 604. By moving an arbitrary thumbnail image 605 displayed in the area 604 to one of the tags 603 by a drag-and-drop method generally performed in a GUI, the true classification result in the lower layer classification and the intermediate layer It is possible to specify both of the true classification results in the classification at the same time. By performing this operation on all the defects, the results to be classified into the intermediate layer and the lower layer are given to all the defects.

  Now, with reference to FIG. 6, the method in which the user teaches the true classification result of the intermediate layer and the lower layer together has been described. However, providing the result to be classified of the intermediate layer complicates the operation. However, user convenience is not good. In order to set the classification criterion while minimizing the burden on the user, it is desirable to allow the user to set the classification criterion only for the lower-layer classification. In the conventionally known automatic classification method, for example, a neural network of a backpropagation type is known as having a multilayer structure, and in this method, only a correct answer of the classification in the final classification of the lower layer is given. It was possible to set the classification criteria for the lower layer and the intermediate layer without teaching the correct answer to the intermediate layer. However, in this method, it is impossible to set an optimal classification standard unless a large number of teaching data are given, and even if the user confirms the result of the intermediate layer, an appropriate classification standard is set. It was impossible to determine if they were. For this reason, applying the conventional neural network method represented by the conventional back propagation is not suitable for applications such as semiconductor defect classification in which a user needs to set a classification criterion for each necessary step. Practical application was difficult.

  To solve this problem, as described in Patent Document 3, it is becoming common to use a rule type as a classification of an intermediate layer. In this method, the determination of the intermediate layer is often represented by a logical expression as shown in, for example, Patent Document 2, and for example, as a criterion for separating concave, convex, and flat described in the publication, A discriminant using only one determined or a small number of feature amounts is reset. That is, while the conventional backpropagation type neural network does not have a specific limitation as an intermediate layer between the result of each intermediate layer and the input, while the intermediate layer is calculated in the rule form, Presence or absence of the relationship between the calculation result of the intermediate layer and the input has already been determined, and this restriction makes it possible to set an appropriate classification criterion from a small number of teacher defects.

  The difficulty in setting the discriminant for the hidden layer is that it is not possible to obtain a correct output for the hidden layer. This problem is solved by evaluating the change in the classification correct rate of the lower layer when the discriminant in the intermediate layer is changed, and setting the discriminant so that the correct rate becomes larger.

When the feature vector used for classifying the intermediate layer of the defect d whose true lower layer classification class is i is Fm, and the feature vector used for the lower layer classification is Fb, the class of the intermediate layer is represented by j. In this case, the discriminant corresponding to the classification class i is, for example, as shown in (Equation 1) below.
f (i) (Fm, Fb)
= Maxj (f1 (j) (Fm) f2 (i, j) (Fb)) (Equation 1)
The necessary conditions for classifying this defect into class i are classified into class j of the intermediate layer where f2 (i, j) (Fb)> f2 (k, j) (Fb) holds for an arbitrary class k. Is Rukoto. Therefore, when the class k is classified into a class j in which a class k that does not satisfy f2 (i, j) (Fb)> f2 (k, j) (Fb) is evaluated, a costly evaluation expression is evaluated. The classification criterion of the intermediate layer is changed by setting the discriminant f1 (j) (Fm) of the intermediate layer so as to minimize it. Since the classification of the intermediate layer is of the rule type, f1 (j) (Fm) is expressed by the following (Equation 2) as a more specific example.

f1 (j) (Fm)
= (Fm1>th1∩Fm2> th2) ∪Fm3> th3 (Equation 2)
By making such a change, it is not always possible to obtain the maximum correct answer rate, but it is possible to change the logical expression so that at least the correct answer rate becomes better than the initial state.

  Unlike a conventionally known method such as a back-propagation type neural network, the change of the discriminant in the method of the present invention involves, for example, changing three variables in the case of the logical expression shown in the above (Equation 2). Therefore, the degree of freedom is small, and more stable performance can be obtained than the teacher defect. On the other hand, since a logical expression is formed based on knowledge obtained in advance, there is an advantage that it is possible to give a complicated expression which is extremely difficult in general learning.

  Although the classification method has been described above, the method of the present invention can of course be implemented in various devices, and one example is a review SEM for reviewing semiconductor defects. In FIG. 7, reference numeral 701 denotes an electron beam source, and 702 denotes a condenser lens. Numerals 703 and 704 denote deflection means in the X and Y directions, respectively, which control the irradiation position of the electron beam in the XY direction. The electron beam deflected by the deflecting means is focused by an objective lens 705 and irradiates a semiconductor wafer 707 as an observation target fixed on an XY stage constituted by 707 and 708. Secondary electrons are generated in the observation target by the electrons applied to the observation target, and the secondary electrons are detected by the secondary electron detection means 709 via the 724 ExB deflector.

  710 and 711 are backscattered electron detectors that detect backscattered electrons from different directions. Reference numeral 716 denotes a unit for imaging the detected electrons. A / D converters 712, 713, and 714 convert the electronic signals detected by 709, 710, and 711 into digital data, and store the digital data in the memory 715. By controlling 703 and 704 to scan the irradiation position of the electron beam two-dimensionally, a two-dimensional image of the secondary electrons detected at 709 and a two-dimensional image of the reflected electrons detected at 710 are stored in the image memory 715. , 711 are formed. The image stored in the unit 716 is processed by the image processing unit 721, and the defect is automatically classified. Reference numeral 717 denotes a defect position specifying unit for calculating a defect area from an image.

  A defect feature extraction unit 718 extracts a defect image feature amount from the two-dimensional defect image formed in the image memory 715. The shape of the defect area calculated in 717 and the brightness in the defect area are used as defect feature amounts. The image feature quantity extracted in 717 and the two-dimensional defect image stored in 716 can be stored in the secondary storage unit 723. Reference numeral 720 denotes a classification criterion setting unit that implements the classification criterion setting method described above. Reference numeral 719 denotes a classifying unit, which classifies defects using the classification standard set by the classification standard setting unit 720. The classifying unit 719 can classify the image features obtained by the image processing unit 717 in real time and classify defects stored in the secondary storage unit 723. Reference numeral 722 denotes a graphic terminal which displays the image stored in 716, the result of classification in 721, or the classification criteria set in 720 to the user. Further, the GUI described with reference to FIG. 6 can also be displayed on this terminal, so that the user can manually teach the classification result. Further, the relationship between the number of teaching defects and the classification accuracy rate described with reference to FIG. 3 can also be displayed here.

  Although an example in which the present invention is implemented in a review SEM has been described here, the present method can be applied to an optical review apparatus or an automatic defect classifier implemented in an inspection apparatus.

4 is a basic sequence according to one embodiment of the present invention. FIG. 4 is an explanatory diagram of a classification criterion setting sequence according to an embodiment of the present invention. It is a basic characteristic explanatory view of a classification standard evaluated when performing a classification standard setting sequence of the present invention. 4 is a sequence of changing a classification criterion in one embodiment of the present invention. 5 is a sequence for selecting a defect to be referred to when setting a classification criterion in one embodiment of the present invention. It is a teaching screen of a classification standard in one example of the present invention. FIG. 4 is an explanatory diagram of a mounting method when the method of the present invention is mounted on a review SEM. FIG. 4 is an explanatory diagram of classification reference reset timing in the defect classification method according to one embodiment of the present invention.

Explanation of reference numerals

201: division number setting step 202: teacher defect division step 203: classification criterion setting step 204: defect classification step not set as a classification criterion 205: classification criterion evaluation result Display step 206: Step of setting classification criteria

Claims (11)

Classifying an image of the first defect group detected by imaging the sample for each defect classification class;
Creating a classification criterion using information of the image of the first defect group classified for each classification class;
Classifying an image of a second defect group on the specimen whose classification class is known based on the created classification criterion;
Obtaining a classification correct rate of the classification criterion based on a result of classifying the image of the second defect group whose classification class is known based on the classification criterion;
Correcting the classification criterion based on the classification accuracy rate data.   In the step of creating the classification criterion, the classification criterion is set using a classifier having a multilayer structure including an intermediate layer and a lower layer, and the classifier of the intermediate layer determines a classification by a logical expression, and the logical expression is Automatically changing the logical formula so that the accuracy of comprehensive classification based on the classifier of the intermediate layer and the classifier of the lower layer is increased. 1. The defect classification method according to 1.   2. The defect classification method according to claim 1, further comprising the step of outputting information relating to the classification accuracy rate obtained in the step of obtaining the classification accuracy rate of the classification criterion.   2. The defect classification method according to claim 1, further comprising the step of: classifying a defect on the sample whose classification class is unknown based on the modified classification criterion in the step of modifying the classification criterion.   5. The defect classification method according to claim 4, wherein the defect whose classification class is unknown is detected in advance by an inspection device, and the positional information on the sample is a known defect.   5. The defect classification method according to claim 4, further comprising the step of resetting the classification criterion corrected in the correcting step based on a result of the classification in the step of classifying the unknown defect. Classifying a group of defects detected on the sample for each defect classification class to generate first teacher data;
Creating a classification criterion using the information of the first teacher data;
Classifying a part of the first teacher data based on the classification criteria;
Obtaining a classification performance of the classification criterion from the result of classification based on the classification criterion,
Creating a second teacher data by deleting from the first teacher data a defect that degrades the classification performance from the first teacher data; and a classification class using the second teacher data. Classifying unknown defects.   In the step of obtaining the classification performance of the classification criterion, a part of the first teacher data with respect to a result of the classification performed in the step of creating the first teacher data for a part of the first teacher data is defined as the classification criterion. 8. The defect classification method according to claim 7, wherein the decomposition performance is obtained based on the accuracy rate of the result of classification based on the classification.   The step of creating the second teacher data includes creating new teacher data by excluding a specific defect from the first teacher data, and using a new classification criterion by using information of the new teacher data. Is generated, a defect of a part of the first teacher data is classified based on the new teacher data, a classification performance of the new classification criterion is obtained from a result of the classification, and the obtained new classification criterion is obtained. By comparing the classification performance of the classification criterion created by using the first teacher data with the classification performance of the classification criterion created by using the first teacher data. 8. The defect classification method according to claim 7, wherein whether the defect is a defect to be classified is evaluated, and when the defect is evaluated as a defect that degrades classification performance, the new data is used as second teacher data. Classification criteria are created using teacher data whose defect classification class is known,
Classification class of the defect classifies the known data based on the classification criteria,
The classification class of the defect determines the classification performance of the classification standard from the result of classifying the known data based on the classification standard,
A method of classifying a defect that corrects the classification criterion based on the obtained classification performance,
In the step of creating the classification criterion, the classification criterion is set using a classifier having a multilayer structure including an upper layer, an intermediate layer, and a lower layer, and the classifier of the intermediate layer is determined to be classified by a logical expression. A defect classification method, wherein a logical expression is set such that a classification correct rate by the lower classifier is minimized when the logical expression is changed.   In the step of determining the classification performance of the classification criterion, the classification performance of the defect is determined based on the correct answer rate for the data of the known classification class as a result of the classification performed in the step of classifying the known data based on the classification standard. The defect classification method according to claim 10, wherein?

Images