Defect classification with optimized purity

• US 8,315,453 B2
• Filed: 07/27/2010
• Issued: 11/20/2012
• Est. Priority Date: 07/27/2010
• Status: Active Grant

• Current Assignee: Applied Materials Israel Limited (Applied Materials Incorporated)
• Original Assignee: Applied Materials Israel Limited (Applied Materials Incorporated)
• Inventors: Shlain, Vladimir, Glazer, Assaf
• Primary Examiner(s): COUSO, JOSE L
• Application Number:
  US12/844,724
• Publication Number:
  US 20120027285A1
• Time in Patent Office: 847 Days
• Field of Search: 382141-152, 382155-159, 382/181, 382/190, 382/209, 382217-222, 382224-228, 382/309, 348 86- 88, 348/92, 348125-130, 706/15, 706/20, 706/900
• US Class Current: 382/149
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/2178 based on feedback of a supervisor
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06F 18/2431 Multiple classes
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06V 30/2504 Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches

Abstract

A method for defect analysis includes identifying single-class classifiers for a plurality of defect classes, the plurality of defect classes characterized by respective ranges of inspection parameter values. Each single-class classifier is configured for a respective class to identify defects belonging to the respective class based on the inspection parameter values, while identifying the defects not in the respective class as unknown defects. A multi-class classifier is identified that is configured to assign each defect to one of the plurality of the defect classes based on the inspection parameter values. Inspection data is received, and both the single-class and multi-class classifiers are applied to the inspection data to assign the defect to one of the defect classes.

Claims

1. A method for defect analysis, comprising:

identifying, by a computer system, respective single-class classifiers for a plurality of defect classes, the plurality of defect classes characterized by respective ranges of inspection parameter values, each single-class classifier being configured for a respective class to identify defects belonging to the respective class based on the inspection parameter values, while identifying the defects not in the respective class as unknown defects;

identifying, by the computer system, a multi-class classifier configured to assign each defect to one of the plurality of the defect classes based on the inspection parameter values;

receiving, by the computer system inspection data with respect to a defect found in a sample; and

automatically applying both the single-class and multi-class classifiers to the inspection data, using the computer system, to assign the defect to one of the defect classes.

2. The method according to claim 1, wherein the multi-class classifier is configured to identify the defects in an overlap region between the respective ranges of at least two of the defect classes as non-decidable defects.

3. The method according to claim 2, and comprising rejecting from classification the unknown defects and the non-decidable defects.

4. The method according to claim 2, wherein specifying the respective single-class and multi-class classifiers comprises jointly setting respective confidence thresholds for distinguishing between known and unknown defects and between decidable and non-decidable defects.

5. The method according to claim 4, wherein setting the respective confidence thresholds comprises adjusting the confidence thresholds to achieve a specified purity target in classification of the defects.

6. The method according to claim 1, wherein automatically applying both the single-class and multi-class classifiers comprises classifying the defect in a given defect class using the multi-class classifier, and verifying that the defect is classified as a known defect by the single-class classifier for the given defect class.

7. The method according to claim 1, wherein specifying the multi-class classifier comprises defining a plurality of decision planes for distinguishing between the respective ranges of the parameter values in a parameter hyperspace for different pairs of the defect classes, and wherein applying the multi-class classifier comprises aggregating multiple binary decisions made with respect to the decision planes.

8. The method according to claim 7, wherein defining the decision planes comprises, for each pair of the defect classes, applying a binary linear support vector machine in order to find a plane giving a maximal separation of the classes in the pair.

9. The method according to claim 1, wherein specifying the respective single-class classifiers comprises defining a one-class support vector machine to distinguish between the defects in a given class and the defects that are classified as unknown defects with respect to the given class.

10. The method according to claim 9, wherein defining the one-class support vector machine comprises mapping the inspection data to a hypersphere in a parameter hyperspace, and finding a hyperplane that cuts the hypersphere to define a hyperspherical cap containing the defects to be classified as belonging to the given class.

11. The method according to claim 1, wherein specifying the respective single-class and multi-class classifiers comprises applying the single-class and multi-class classifiers to a set of training data in order to define the respective ranges of the inspection parameter values for use in applying the single-class and multi-class classifiers to the inspection data.

12. A method for analyzing defects, comprising:

defining, by a computer system, a plurality of defect classes;

receiving, by the computer system, a training set comprising inspection data with respect to defects that have been classified as belonging to respective defect classes;

training, by the computer system, a plurality of classifiers using the training set to define a respective range of inspection parameter values that characterizes each defect class and, using the parameter values, to classify each defect as belonging to a respective class with a respective level of confidence;

adjusting, by the computer system, a confidence threshold for rejection of defect classifications having low levels of confidence to achieve a specified purity target in classification of the training set; and

applying, by the computer system, the trained classifiers with the adjusted confidence threshold to further inspection data outside the training set,

wherein the plurality of classifiers comprises;

a multi-class classifier, having a first confidence threshold such that the defects classified by the multi-class classifier as falling below the first confidence threshold are identified as non-decidable defects; and

at least one single-class classifier, having a second confidence threshold such that defects classified by the at least one single-class classifier as falling below the second confidence threshold are identified as unknown defects.

13. The method according to claim 12, wherein training the one or more computerized classifiers comprises training a plurality of classifiers, having respective confidence thresholds that are jointly adjusted to balance a purity of the classification against a rejection rate.

14. The method according to claim 13, wherein adjusting the confidence threshold comprises maximizing the purity of the classification of the training set while keeping the rejection rate no greater than a predetermined maximum.

15. The method according to claim 12, wherein training the one or more computerized classifiers comprises:

dividing the training set into multiple folds;

deriving the range of inspection parameter values based on a training subset of the folds;

testing the derived range on a testing subset of the folds, disjoint from the training subset; and

repeating the steps of deriving and testing over multiple, different training and

testing subsets of the folds.

16. Apparatus for defect analysis, comprising:

a memory, configured to store respective ranges of inspection parameter values for a plurality of defect classes; and

a processor, which is configured to receive inspection data with respect to a defect found in a sample, and to apply both single-class and multi-class classifiers to the inspection data, based on the inspection parameter values, to assign the defect to one of the defect classes,

wherein each single-class classifier is configured for a respective class to identify defects belonging to the respective class, while identifying the defects not in the respective class as unknown defects, and the multi-class classifier is configured to assign each defect to one of the plurality of the defect classes.

17. The apparatus according to claim 16, wherein the multi-class classifier is configured to identify the defects in an overlap region between the respective ranges of at least two of the defect classes as non-decidable defects.

18. The apparatus according to claim 17, wherein the processor is configured to reject from classification the unknown defects and the non-decidable defects.

19. The apparatus according to claim 17, wherein confidence thresholds for distinguishing between known and unknown defects and between decidable and non-decidable defects are set jointly for the single-class and multi-class classifiers.

20. The apparatus according to claim 19, wherein the confidence thresholds are adjusted to achieve a specified purity target in classification of the defects.

21. The apparatus according to claim 16, wherein the processor is configured to classify the defect in a given defect class using the multi-class classifier, and to verify that the defect is classified as a known defect by the single-class classifier for the given defect class.

22. The apparatus according to claim 16, wherein the multi-class classifier is configured to apply a plurality of decision planes for distinguishing between the respective ranges of the parameter values in a parameter hyperspace for different pairs of the defect classes, and to aggregate multiple binary decisions made with respect to the decision planes.

23. The apparatus according to claim 22, wherein the decision planes are defined, for each pair of the defect classes, by applying a binary linear support vector machine in order to find a plane giving a maximal separation of the classes in the pair.

24. The apparatus according to claim 16, wherein each of the single-class classifiers is configured to apply a one-class support vector machine to distinguish between the defects in a given class and the defects that are classified as unknown defects with respect to the given class.

25. The apparatus according to claim 24, wherein the one-class support vector machine is operative to map the inspection data to a hypersphere in a parameter hyperspace, and to find a hyperplane that cuts the hypersphere to define a hyperspherical cap containing the defects to be classified as belonging to the given class.

26. The apparatus according to claim 16, wherein the respective ranges of the inspection parameter values for the single-class and multi-class classifiers are defined by applying the single-class and multi-class classifiers to a set of training data.

27. Apparatus for analyzing defects, comprising:

a memory, which is configured to store information with respect to a plurality of defect classes; and

a processor, which is configured to receive a training set comprising inspection data with respect to defects that have been classified as belonging to respective defect classes, to train a plurality of classifiers using the training set to define a respective range of inspection parameter values that characterizes each defect class and, using the parameter values, to classify each defect as belonging to a respective class with a respective level of confidence, to adjust a confidence threshold for rejection of defect classifications having low levels of confidence to achieve a specified purity target in classification of the training set, and to apply the trained classifiers with the adjusted confidence threshold to further inspection data outside the training set, wherein the plurality of classifiers comprises;

a multi-class classifier, having a first confidence threshold such that the defects classified by the multi-class classifier as falling below the first confidence threshold are identified as non-decidable defects; and

at least one single-class classifier, having a second confidence threshold such that defects classified by the at least one single-class classifier as falling below the second confidence threshold are identified as unknown defects.

28. The apparatus according to claim 27, wherein the one or more classifiers comprise a plurality of classifiers, having respective confidence thresholds that are jointly adjusted to balance a purity of the classification against a rejection rate.

29. The apparatus according to claim 28, wherein the processor is configured to adjust the confidence threshold to maximize the purity of the classification of the training set while keeping the rejection rate no greater than a predetermined maximum.

30. The apparatus according to claim 27, wherein the processor is configured to train the one or more classifiers by dividing the training set into multiple folds, deriving the range of inspection parameter values based on a training subset of the folds, testing the derived range on a testing subset of the folds, disjoint from the training subset, and repeating the steps of deriving and testing over multiple, different training and testing subsets of the folds.

31. A computer readable storage medium including instructions that, when executed by a processing device, cause the processing device to implement a method for defect analysis, the method comprising:

identifying, by a computer system, respective single-class classifiers for a plurality of defect classes, the plurality of defect classes characterized by respective ranges of inspection parameter values, each single-class classifier being configured for a respective class to identify defects belonging to the respective class based on the inspection parameter values, while identifying the defects not in the respective class as unknown defects;

identifying, by the computer system, a multi-class classifier configured to assign each defect to one of the plurality of the defect classes based on the inspection parameter values;

receiving, by the computer system inspection data with respect to a defect found in a sample; and

automatically applying both the single-class and multi-class classifiers to the inspection data, using the computer system, to assign the defect to one of the defect classes.

32. The computer readable storage medium according to claim 31, wherein the multi-class classifier is configured to identify the defects in an overlap region between the respective ranges of at least two of the defect classes as non-decidable defects.

33. The computer readable storage medium according to claim 32, the method further comprising rejecting from classification the unknown defects and the non-decidable defects.

34. The computer readable storage medium according to claim 32, wherein confidence thresholds for distinguishing between known and unknown defects and between decidable and non-decidable defects are set jointly for the single-class and multi-class classifiers.

35. The computer readable storage medium according to claim 34, wherein the confidence thresholds are adjusted to achieve a specified purity target in classification of the defects.

36. The computer readable storage medium according to claim 31, wherein the method further comprising classifying the defect in a given defect class using the multi-class classifier, and verifying that the defect is classified as a known defect by the single-class classifier for the given defect class.

37. The computer readable storage medium according to claim 31, wherein the multi-class classifier is configured to apply a plurality of decision planes for distinguishing between the respective ranges of the parameter values in a parameter hyperspace for different pairs of the defect classes, and to aggregate multiple binary decisions made with respect to the decision planes.

38. The computer readable storage medium according to claim 37, wherein the decision planes are defined, for each pair of the defect classes, by applying a binary linear support vector machine in order to find a plane giving a maximal separation of the classes in the pair.

39. The computer readable storage medium according to claim 31, wherein each of the single-class classifiers is configured to apply a one-class support vector machine to distinguish between the defects in a given class and the defects that are classified as unknown defects with respect to the given class.

40. The computer readable storage medium according to claim 39, wherein the one-class support vector machine is operative to map the inspection data to a hypersphere in a parameter hyperspace, and to find a hyperplane that cuts the hypersphere to define a hyperspherical cap containing the defects to be classified as belonging to the given class.

41. The computer readable storage medium according to claim 31, wherein the respective ranges of the inspection parameter values for the single-class and multi-class classifiers are defined by applying the single-class and multi-class classifiers to a set of training data.

42. A computer readable storage medium including instructions that, when executed by a processing device, cause the processing device to implement a method for defect analysis, the method comprising:

defining, by a computer system, a plurality of defect classes;

receiving, by the computer system, a training set comprising inspection data with respect to defects that have been classified as belonging to respective defect classes;

training, by the computer system, one or more computerized classifiers using the training set to define a respective range of inspection parameter values that characterizes each defect class and, using the parameter values, to classify each defect as belonging to a respective class with a respective level of confidence;

adjusting, by the computer system, a confidence threshold for rejection of defect classifications having low levels of confidence to achieve a specified purity target in classification of the training set; and

applying, by the computer system, the trained classifiers with the adjusted confidence threshold to further inspection data outside the training set.

43. The computer readable storage medium according to claim 42, wherein the one or more classifiers comprise a plurality of classifiers, having respective confidence thresholds that are jointly adjusted to balance a purity of the classification against a rejection rate.

44. The computer readable storage medium according to claim 43, wherein the plurality of the classifiers comprise:

a multi-class classifier, having a first confidence threshold such that the defects classified by the multi-class classifier as falling below the first confidence threshold are identified as non-decidable defects; and

at least one single-class classifier, having a second confidence threshold such that defects classified by each single-class classifier as falling below the second confidence threshold are identified as unknown defects.

45. The computer readable storage medium according to claim 43, the method further comprising adjusting the confidence threshold to maximize the purity of the classification of the training set while keeping the rejection rate no greater than a predetermined maximum.

46. The computer readable storage medium according to claim 42, the method further comprising training the one or more classifiers by dividing the training set into multiple folds, deriving the range of inspection parameter values based on a training subset of the folds, testing the derived range on a testing subset of the folds, disjoint from the training subset, and repeating the steps of deriving and testing over multiple, different training and testing subsets of the folds.