Patent Application
20110188735
The invention relates to a technology for inspecting a semiconductor wafer, and more particularly, to an effective technology applied to the method for setting a defect classification criterion of an inspection device.
Miniaturization of the semiconductor has been markedly advanced accompanied with the recent trend of compact and highly sophisticated electronic products, resulting in new products which hit the market in succession. Meanwhile in the semiconductor manufacturing step, the in-line defect inspection of the semiconductor wafer is conducted. Accompanied with miniaturization of the semiconductor, the defect as the cause of the device failure, that is, the defect of interest (DOI) has also been miniaturized. The highly sensitive defect inspection which is capable of coping with such miniaturization has been demanded. Several tens of thousands of defects of disinterest (nuisances) on the wafer such as negligible irregular surface are detected, resulting in the state where a few DOIs exist in a large number of defects which include nuisances.
It is therefore important to ensure detection only of the DOIs of the new device. The method for automatically classifying the defect has been proposed as Auto Defect Classification (ADC) conducted by analyzing the image obtained from the inspection which has been conducted using the appearance tester. Alternatively, a method has been proposed for automatically classifying further detailed images of the defect, which have been detected again after conducting the inspection using the appearance inspection device.
Various methods have been proposed for conducting the ADC, which include a rule type process for classifying the defect features including plural image featured values such as brightness and defect shape extracted from the image into the defect class based on the predetermined rule, an instruction type for setting plural scalar values each as a group of the respective items of the defect features to a multidimensional vector to automatically generate the criterion for classifying the detects based on distribution of the defect class in the multidimensional space formed by the multidimensional vector, and further a combination type formed by combining the rule type and the instruction type.
In order to automatically classify the defect by conducting the ADC, the defect classification criterion is required to be set before automatic classification based on the defect sample with a known classification class.
In the case of using the rule type, generally, it is necessary to set determination threshold values corresponding to some items of the defect feature. In the case of using the instruction type, it is necessary to obtain distribution of the defect class in the multidimensional space. In the state where a large number of defects are detected by the inspection device, the method which allows appropriate and easy setting of the defect classification criterion is indispensable.
As for setting of the defect classification criterion, Patent Document 1 discloses the image recognition device for recognizing the classification class of the online image by comparing the sample image data with normal image data to obtain the appropriate defect classification pattern to be instructed, and further obtaining the featured value with respect to the defects. Patent. Document 2 discloses the inspection device structured to evaluate the classification criterion when right or wrong of classification of the defect group with a known classification class based on the preliminarily obtained classification criterion to the known value is relatively low. Patent Document 3 discloses the automatic classification device provided with the function for updating the instruction data used for automatic classification based on the defect image information on the basis of the feature of the defect image.
With the method disclosed in Patent Document 1, the user instructs the instruction data found from the defect image data. The user also instructs the data required to be corrected among the classified defect image data. In any of the cases, it is up to the user to select the defect image.
The method disclosed in Patent Document 2 discloses the classification criterion which becomes more stable as the increase in the instruction data. However, the classification criterion which may be derived from less of instruction data is not disclosed.
With the method disclosed in Patent Document 3, the user instructs the defect class of each of the collected defect images. In this case, at least one of the existing instruction data and the newly collected defect image data may be used for generating the new instruction data through generally employed method for generating instruction data.
In order to ensure detection of the DOI, it is necessary to make sure to instruct the DOI. However, appropriate instruction is not easy in the state where a few DOIs exist in a large number of nuisances. For this, the user may be forced to bear the burden of instructing the defect while confirming several tens of defects one by one. Otherwise, as a result of instructing only some of the defects, the classification criterion cannot be optimized, resulting in missing of the DOI or misinformation where the nuisance is incorrectly classified as DOI.
It is a first object of the present invention to provide a method and a device for inspection capable of improving the classification performance by a few appropriate defect instructions even in the state where a few DOIs exist in a large number of nuisances during the defect inspection.
It is a second object of the present invention to provide a method and a device for inspection capable of ensuring high classification performance while mitigating the burden of the user's defect instructions even in the state where a few DOIs exist in a large number of nuisances during the defect inspection.
For the purpose of achieving the aforementioned objects, the present invention provides an inspection method including: a defect extraction step of extracting one or more defects from plural defects detected by imaging a sample; a defect image display step of displaying an image of the extracted defect; a defect classification class input step of inputting a classification class of the displayed defect; a classification criterion calculation step of calculating a classification criterion from image information and classification class of the defects which have been extracted; a classification performance determination step of determining a performance of the defect classification based on the classification criterion; and an inspection step of inspecting unknown defects based on the classification criterion calculated in the classification criterion calculation step.
For the purpose of achieving the aforementioned objects, the present invention provides an inspection device including: defect extraction means for extracting one or more defects from plural defects detected by imaging a sample; defect image display means for displaying an image of the extracted defect; defect classification class input means for inputting a classification class of the displayed defect; classification criterion calculation means for calculating a classification criterion from image information and classification class of the defects which have been extracted; classification performance determination means for determining a performance of the defect classification based on the classification criterion; and inspection means for inspecting unknown defects based on the classification criterion calculated by the classification criterion calculation means.
The present invention provides the method and device for inspection capable of improving the classification performance by a few appropriate defect instructions even in the state where a few DOIs exist in a large number of nuisances in the defect inspection.
Furthermore, the preset invention provides the method and device for inspection capable of ensuring high classification performance while mitigating the burden of the user's defect instructions even in the state where a few DOIs exist in a large number of nuisances in the defect inspection.
[
[
[
[
[
[
[
[
[
[
A first embodiment of the present invention will be described referring to the drawings.
The determination condition setting unit 612 stores conditions based on which a determination is made with respect to the defect of the semiconductor wafer. The image processing circuit 610 processes the image, and makes the defect determination based on the condition so as to extract the defect image. The defect image is transmitted to the classification condition setting unit 500 via the general control unit 613. The classification condition setting unit 500 includes an image processing unit 502 which processes the image of the defect to extract the featured value, a defect classification unit 503 which extracts the defect by calculating the featured value, creates the classification criterion, and calculates the classification performance, a data storage unit 506 which stores the classification criterion, the defect image, the defect featured value and the defect classification, and a user interface unit 507 which displays the defect image and the defect featured value on the screen, and allows the user to input the defect classification instruction. They are connected with one another so that data is transmitted and received as necessary.
The aforementioned means will be described in accordance with the embodiment of the present invention. First of all, the subject wafer is selected, and an instruction is conducted.
Clicking an instruction tab 206 allows transition to the instruction screen. (foreign substance defect), and ⋄ (open defect) each indicating the defect type. Classification criteria 330 to 332 (see
An image 301 of the automatically extracted defect is displayed at the upper right section on the, screen. A classification class is displayed at the right side of the image. When a classification class input column 310 is designated, the subject defect is displayed. Upon selection of the defect by the user, the defect type may be designated.
An accuracy rate table 312 indicating right or wrong is displayed at the lower section, the X-axis of which indicates the defect type determined by the aforementioned means and the Y-axis of which indicates the defect type instructed by the user using the classification class input column 310. Referring to the short-circuit column, the means indicates that the user has instructed three short-circuit defects. With the means, the grain defect is determined one time. As a result, the accuracy rate becomes ⅔=67%. A graph 313 representing transition of classification performance is displayed at the lower right side.
It is assumed that a large number of defects detected from the selected wafer have the featured values calculated using the generally employed method. The respective defects are classified into respective clusters using a known multilevel clustering method.
When two featured values are selected by the first featured value designation button 305 and the second featured value designation button 306, locations of the clusters are displayed with such codes as ◯ and Δ on the featured value space map 302.
Depending on the number of the cluster types, one or more defects to be instructed are automatically extracted through the predetermined process. The automatically extracted defect image 301 is sequentially displayed on the screen.
The correct classification class of displayed images of 10 defects are instructed in the classification class input column 310. The instructed content may be different from that of the cluster classified using the generally employed method.
The classification criterion and classification performance are calculated using the instructed defect classification class and the featured value information.
The initial classification criterion is calculated using the neural network method as disclosed in Patent Document 3, for example. As the classification class of the instructed defect and the featured value are known, such information is input into the neural network which weights the featured value with a predetermined weight coefficient. Learning is conducted so that output information derived from the neural network is set to correspond to the defect classification class. In other words, in the learning, the obtained neural network output information and the defect classification class are compared. When an error value indicating disparity state exceeds the predetermined threshold value, the weight coefficient is corrected in accordance with the error value. Then the same defect data is input again to weight the featured value with the corrected weight coefficient. The aforementioned process is repeatedly performed until the error value becomes equal to or smaller than the threshold value.
According to the embodiment, in consideration of the distributed state of 10 defects instructed by the user as illustrated on the featured value space map 302, the featured value space map 302 is divided by the following three lines each as the classification criterion.
a1×f1+b1×f2+c1=0
a2×f1+b2×f2+c2=0
a3×f1+b3×f2+c3=0
All the defects on the featured value space map 302 will be determined based on the calculated classification criterion.
The classification class of each defect is determined in accordance with the following judgment conditions. The featured value space map 302 shown in
If a1×f1i+b1×f2i+c1≧0̂a2×f1i+b2×f2i+c2≧0, the class corresponds to the foreign substance;
If a1×f1i+b1×f2i+c1≧0̂a2×f1i+b2×f2i+c2<0, the class corresponds to the open;
If a1×f1i+b1×f2i+c1<0̂a3×f1i+b3×f2i+c3≧0, the class corresponds to the grain; and
If a1×f1i+b1×f2i+c1<0̂a3×f1i+b3×f2i+c3<0, the class corresponds to the short-circuit.
According to the explanation as described above, three lines are used to define the classification criteria in accordance with the distribution state of 10 defects instructed by the user shown on the featured value space map 302. However, the known method does not have to be used for classifying the cluster so long as it is clear that use of the three lines is capable of defining the classification criterion in accordance with the defect distribution even if the clusters are not classified without conducting the cluster classification.
The other method using circle or semi-circle with a center may be employed in accordance with the distribution state.
Meanwhile, if five of 10 automatically extracted defects by the initial defect presenting means 101 coincide with the content instructed by the user, the classification performance may be calculated, that is, 5/10=50%, which will be displayed on the classification performance transition graph 313 indicating transition of the accuracy rate. The accuracy rate table 312 shows the defect class each automatically extracted and the defect class instructed by the user.
One or more defects to be instructed next are automatically extracted from those detected through inspection, and the image 301 of the automatically extracted defect is displayed on the screen likewise the initial presenting means 101. The automatic extraction of the defect to be instructed is conducted by extracting the defect around the boundary of the clusters, or automatically extracting the defect which is the closest to the gravity center of the adjacent cluster by applying the division optimized clustering method such as well-known k-means method. In this case, six defects 333 to 338 are automatically extracted as shown in
The user instructs the classification class of the defect having its image displayed using the input column 310 like the case of the initial classification class instruction means 102.
Like the initial classification criterion and classification performance calculation means 103, the classification criterion and the classification performance are calculated by executing the predetermined process using the classification class of the instructed defect and the featured value information. In this case, the classification criterion and classification performance are calculated with respect to total of 16 defects including 10 defects automatically extracted by the initial defect instruction means 101 and six defects automatically extracted by the defect presenting means 104 using the method represented by the initial classification class instruction means 102. The respective results are displayed on the featured value space map 302, the classification performance transition graph 313 and the accuracy rate table 312. The featured value space map is corrected likewise the initial classification class instruction means 102 as shown in
The classification performance calculated by the previous classification criterion and classification performance calculation means 106, and the classification performance calculated by the classification criterion and classification performance calculation means 106 earlier than the previous means, or the initial classification criterion and classification performance calculation means 103 are compared. Alternatively, the immediately previous classification performance calculated by the classification criterion and classification performance calculation means 106 and the classification performance calculated by the classification criterion and classification performance calculation means 106 one cycle before may also be compared. When the classification performance one cycle before is of the initial classification criterion and classification performance calculation means 103, it may be subjected to the comparison. As the transition of the classification performance may vary at a small interval, the comparison may be made after calculating the average movement with respect to the classification performance transition. If the calculated classification performance is higher than the one calculated one means before, it is determined that the classification performance is improved. If the calculated classification performance is lower or has hardly changed compared with the performance calculated one cycle before, it is determined that the classification performance has not been improved.
The classification performance transition graph 313 is used herein. The classification performance transition graph 313 plots the classification performance for each of the initial classification criterion and classification performance calculation means 103 and the classification criterion and classification performance calculation means 106 while defining X-axis as the number of operating the classification criterion and classification performance calculation means 106, and Y-axis as the classification performance.
When the classification performance is improved, the process returns to the defect presenting means 104 where one or more defects image from the defects, the classification classes of which have not been instructed is displayed on the screen. Thereafter, the process will be repeatedly executed as described above. If the classification performance is no longer improved, the process proceeds to (8) storage means 108 where the classification criterion obtained in the aforementioned process is stored as the set value. Thereafter, the inspection•classification will be executed using the set classification criterion.
In a stage where the sequence of the criterion setting 401 is finished, the optimal defect classification result of the subject wafer may be obtained. Accordingly, the sequence of the criterion setting 401 may be set as the procedure of the normal inspection method.
In the aforementioned means and sequence, the multilevel clustering process is employed for automatically extracting the defect in the initial defect presenting means 101. However, any other method is applicable for each means, and accordingly, such method may be employed.
The embodiment provides method and device for inspection capable of improving the classification performance by a few appropriate defect instructions even in the state where a few DOIs exist in a large number of nuisances in the defect inspection.
The embodiment provides method and device for inspection capable of ensuring high classification performance while mitigating the burden of the user's defect instructions even in the state where a few DOIs exist in a large number of nuisances in the defect inspection.
The embodiment allows improvement of the classification performance with a few appropriate defect instructions by repeatedly instructing the classification class of the defect image automatically displayed on the screen by the user. This may ensure the high classification performance while mitigating the burden of the user's defect instructions.
According to the embodiment, the sequence of the criterion setting 401 is employed as the procedure of the normal inspection method. In this case, an updating classification criterion value is calculated. If the updating classification criterion value is largely different from the existing criterion value, it is considered that the process or the like has been changing. The means shown in
In the aforementioned description, the classification condition setting unit 500 is integrally formed with the main body of the device. However, the device may be structured to allow the external device to perform setting of the classification criterion value while having the units up to the general control unit 613 required for extracting the defect from the defect images built in the main body of the device. An optical inspection device may be employed as the device employed for the aforementioned case. An example of the optical inspection is illustrated in
The aforementioned optical inspection device is capable of providing the effect of the present invention when it is used together with the external device. Explanation of code
| Number | Date | Country | Kind |
|---|---|---|---|
| 2008-219732 | Aug 2008 | JP | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/JP2009/002549 | 6/5/2009 | WO | 00 | 4/20/2011 |
