Machine learning-based classification of defects in a semiconductor specimen
There is provided a method of automated defects' classification, and a system thereof. The method comprises obtaining data informative of a set of defects' physical attributes usable to distinguish between defects of different classes among the plurality of classes; training a first machine learning model to generate, for the given defect, a multi-label output vector informative of values of the physical attributes, thereby generating for the given defect a multi-label descriptor; and using the trained first machine learning model to generate multi-label descriptors of the defects in the specimen. The method can further comprise obtaining data informative of multi-label data sets, each data set being uniquely indicative of a respective class of the plurality of classes and comprising a unique set of values of the physical attributes; and classifying defects in the specimen by matching respectively generated multi-label descriptors of the defects to the multi-label data sets.
1. A method of automated classifying defects in a semiconductor specimen into a plurality of classes, the method comprising, by a processing and memory circuitry (PMC):
obtaining a set of physical attributes of the defects usable to distinguish between defects of different classes among the plurality of classes;
obtaining data informative of a plurality of multi-label data sets, each multi-label data set of the plurality of multi-label data sets being uniquely indicative of a respective class of the plurality of classes and comprising a unique set of a plurality of respective values of corresponding physical attributes from the set of physical attributes characterizing the respective class, the plurality of respective values serving as a plurality of labels for each defect in the respective class; and
training a first machine learning model, using the data, to process a sample comprising one or more images informative of a given defect so to generate for the given defect a multi-label output vector informative of a plurality of respective values of corresponding physical attributes from the set of physical attributes characterizing the given defect, thereby generating a multi-label descriptor for the given defect.
2. The method of claim 1 further comprising, by the PMC:
training a second machine learning model, using the data, to provide a multi-label classification for the given defect.
3. The method of claim 2 , wherein providing a multi-label classification of the given defect includes defining a certainty threshold as a ratio between a number of values in the multi-label descriptor that match to a given class and a total number of values in the multi-label data set indicative of the given class.
4. The method of claim 3 further comprising using the certainty threshold to enable at least one of:
a. optimizing of confidence levels of defects classification;
b. identifying misclassified defects;
c. setting purity requirements separately for each class and/or group of classes;
d. setting accuracy requirements separately for each class and/or group of classes; and
e. setting extraction requirements separately for each class and/or group of classes.
5. The method of claim 1 , wherein the physical attributes in the set of physical attributes of the defects are informative of at least one of: physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, layer, texture, edges and chemical composition.
6. The method of claim 1 , wherein the plurality of classes comprises a “particle” class and a “bridge” class, and wherein the set of defects' physical attributes comprises roughness of texture, clearness of edges, position in relation to a top of a pattern, and position in relation to two patterns.
7. The method of claim 2 , wherein the values of the physical attributes in the multi-label descriptors and the multi-label data sets are binary.
8. The method of claim 2 , wherein the values of the physical attributes in the multi-label descriptors and the multi-label data sets correspond to “Yes”, “No” and “Not relevant” with regard to respective physical attributes.
9. A system to classify defects in a semiconductor specimen into a plurality of classes, the system comprising a processing and memory circuitry (PMC) operatively connected to an input interface, wherein the input interface is configured to receive samples comprising images informative of the defects; and wherein the PMC is configured to:
use a trained first machine learning model to generate multi-label descriptors of the defects in the semiconductor specimen, the multi-label descriptors being usable for classification,
wherein the first machine learning model is previously trained using data informative of a plurality of multi-label data sets, each multi-label data set of the plurality of multi-label data sets being uniquely indicative of a respective class of the plurality of classes, and comprising a unique set of a plurality of respective values of corresponding physical attributes from a set of physical attributes of the defects characterizing the respective class, the plurality of respective values serving as a plurality of labels of each defect in the respective class, the first machine learning model trained to process a sample comprising one or more images informative of a given defect so to generate for the given defect a multi-label output vector informative of a plurality of respective values of corresponding physical attributes from the set of physical attributes characterizing the given defect, thereby generating a multi-label descriptor for the given defect.
10. The system of claim 9 , wherein the PMC is further configured to:
upon training a second machine learning model to provide a multi-label classification, use the trained second machine learning model for classifying the defects in the semiconductor specimen by matching respectively generated multi-label descriptors of the defects to the multi-label data sets.
11. The system of claim 9 , wherein the PMC is further configured to:
analyze the generated multi-label descriptors of the defects in the semiconductor specimen to recognize new repeating multi-label data sets, thereby identifying new classes of the defects.
12. The system of claim 9 , wherein the PMC is further configured to recognize multi-modal behavior of one or more classes by analyzing the generated multi-label descriptors of the defects in the specimen.
13. The system of claim 10 , wherein classifying a defect of the defects in the semiconductor specimen includes defining a certainty threshold as a ratio between a number of values in the multi-label descriptor that match to a given class and a total number of values in the multi-label data set indicative of the given class.
14. The system of claim 13 , wherein the PMC is further configured to use the certainty threshold to enable at least one of:
a. optimizing of confidence levels of defects classification;
b. identifying misclassified defects;
c. setting purity requirements separately for each class and/or group of classes;
d. setting accuracy requirements separately for each class and/or group of classes; and
e. setting extraction requirements separately for each class and/or group of classes.
15. The system of claim 9 , wherein the physical attributes in the set of physical attributes of the defects are informative of at least one of: physical location, shape, perimeter, sidewall angle, aspect ratio, orientation, symmetry, layer, texture, edges and chemical composition.
16. The system of claim 9 , wherein the plurality of classes comprises a “particle” class and a “bridge” class, and wherein the set of defects' physical attributes comprises roughness of texture, clearness of edges, position in relation to a top of a pattern, and position in relation to two patterns.
17. A non-transitory computer readable medium comprising instructions that, when executed by a computer, cause the computer to perform a method of classifying defects in a semiconductor specimen into a plurality of classes, the method comprising:
using a trained first machine learning model to generate multi-label descriptors of the defects in the semiconductor specimen, the multi-label descriptors being usable for classification,
wherein the first machine learning model is previously trained using data informative of a plurality of multi-label data sets, each multi-label data set of the plurality of multi-label data sets being uniquely indicative of a respective class of the plurality of classes, and comprising a unique set of a plurality of respective values of corresponding physical attributes from a set of physical attributes of the defects characterizing the respective class, the plurality of respective values serving as a plurality of labels of each defect in the respective class, the first machine learning model trained to process a sample comprising one or more images informative of a given defect so to generate for the given defect a multi-label output vector informative of a plurality of respective values of corresponding physical attributes from the set of physical attributes characterizing the given defect, thereby generating a multi-label descriptor for the given defect.
18. The non-transitory computer readable medium of claim 17 , wherein the method further comprises:
upon training a second machine learning model to provide a multi-label classification, using the trained second machine learning model to classify the defects in the semiconductor specimen by matching respectively generated multi-label descriptors of the defects to the multi-label data sets.