Machine learning method and apparatus for inspecting reticles
Apparatus and methods for inspecting a photolithographic reticle are disclosed. A reticle inspection tool is used at one or more operating modes to obtain images of a plurality of training regions of a reticle, and the training regions are identified as defect-free. Three or more basis training images are derived from the images of the training regions. A classifier is formed based on the three or more basis training images. The inspection system is used at the one or more operating modes to obtain images of a plurality of test regions of a reticle. Three or more basis test images are derived from to the test regions. The classifier is applied to the three or more basis test images to find defects in the test regions.
1. A method of inspecting a photolithographic reticle, the method comprising:
using an inspection system at one or more operating modes, obtaining images of a plurality of training regions of a reticle, wherein the training regions are identified as defect-free;
deriving three or more basis training images from the images of the training regions;
forming a classifier based on the three or more basis training images by:
mapping the three or more basis training images to a plurality of voxels of a three or more dimensional space; and
assigning an observation indicator to each voxel so as to specify whether such voxel is defined as a non-defective voxel that is observed in the training set;
using the inspection system at the one or more operating modes, obtaining images of a plurality of test regions of a reticle; and
deriving three or more basis test images from the test regions;
and
applying the classifier to the three or more basis test images to find defects in the test regions.
2. The method of claim 1 , wherein forming the classifier further includes:
healing a plurality of holes in the non-defective voxels by changing an indicator of each voxel of each hole from specifying such voxel is not observed in the training set to specifying that such voxel is a non-defective voxel.
3. The method of claim 2 , wherein forming the classifier further includes performing a distance transformation to assign a distance value to each voxel based on such voxel's distance to a non-defective voxel.
4. The method of claim 3 , wherein a defect is found when a particular voxel as a distance value that is above a predefined threshold.
5. The method of claim 1 , further comprising:
forming each of a plurality of classifiers based on a plurality of different sets of the three or more basis training images; and
applying the plurality of classifiers to the three or more basis test images to find defects in the test regions.
6. The method of claim 1 , further comprising forming a space-partitioning data structure from the classifier.
7. The method of claim 1 , wherein forming the classifier comprises approximating a set of feature vectors that correspond to the non-defective voxels with a collection of subsets of affine-linear spaces.
8. The method of claim 1 , wherein forming the classifier comprises:
clustering a set of feature vectors that are based on the basis training images into a plurality of clusters;
determining a cluster center of each cluster's feature vectors;
for each cluster, determining a plurality of cluster vectors from each cluster center to the cluster's feature vectors; and
for each cluster, determining an affine-linear space that includes the cluster's cluster center and that is spanned by a plurality of principal components of the cluster's cluster vectors,
wherein applying the classifier to the three or more basis test images is based on a distance between each test vector of the three or more basis test images and a nearest affine-linear space.
9. The method of claim 1 , wherein forming the classifier comprises:
minimizing a norm of distances from a affine linear space to a plurality of feature vectors from the three or more basis training images;
discarding a set of the feature vectors for which a distance is highest; and
repeating the step for minimizing the norm of distance to refine a plurality of affine-linear spaces to better fit remaining ones of the feature vectors until the feature vectors are fitted to a final set of the affine-linear spaces within a predefined tolerance level.
10. The method of claim 1 , wherein the classifier includes three or more dimensions of feature vectors that each has a transmitting image value, a reflecting image value, and a transformed image value based on transformation of a transmitting or reflecting image value.
11. The method of claim 10 , wherein each transformed image value is determined by one of the following: ∥∇T∥, ∥∇R∥, ∇ 2 T, ∇ 2 R, wherein T is a transmitting image value and R is a reflecting image value.
12. The method of claim 1 , wherein the one or more operating modes are selectable to differ in one or more of the following qualities or quantities: a reflecting mode, a transmitting mode, a brightfield mode, a darkfield mode, an illumination numerical aperture, an illumination pupil pattern, an illumination polarization, a collection numerical aperture, a collection pupil filter amplitude and phase, a collection analyzer, and a focus.
13. An inspection system for inspecting a photolithographic reticle, the system comprising at least one memory and at least one processor that are configured to perform the following operations:
at one or more operating modes of the inspection system, obtaining images of a plurality of training regions of a reticle, wherein the training regions are identified as detect-free;
deriving three or more basis training images from the images of the training regions;
forming a classifier based on the three or more basis training images by:
mapping the three or more basis training images to a plurality of voxels of a three or more dimensional space; and
assigning an observation indicator to each voxel so as to specify whether such voxel is defined as a non-defective voxel that is observed in the training set;
at the one or more operating modes, Obtaining images of a plurality of test regions of a reticle; and
deriving three or more basis test images from the test regions; and
applying the classifier to the three or more basis test images to find defects in the test regions.
14. The inspection system of claim 13 , wherein forming the classifier further includes:
healing a plurality of holes in the non-defective voxels by changing an indicator of each voxel of each hole from specifying such voxel is not observed in the training set to specifying that such voxel is a non-defective voxel.
15. The inspection system of claim 14 , wherein forming the classifier further includes performing a distance transformation to assign a distance value to each voxel based on such voxel's distance to a non-defective voxel.
16. The inspection system of claim 15 , wherein a defect is found when a particular voxel has a distance value that is above a predefined threshold.
17. The inspection system of claim 13 , wherein the at least one memory and at least one processor are further configured for:
forming each of a plurality of classifiers based on a plurality of different sets of the three or more basis training images; and
applying the plurality of classifiers to the three or more basis test images to find defects in the test regions.
18. The inspection system of claim 13 , wherein the at least one memory and at least one processor are further configured for forming a space-partitioning data structure from the classifier.
19. The inspection system of claim 13 , herein forming the classifier comprises approximating a set of feature vectors that correspond to the non-defective voxels with a collection of subsets of affine-linear spaces.
20. The inspection system of claim 13 , wherein forming the classifier comprises:
clustering a set of feature vectors that are based on the basis training images into a plurality of clusters;
determining a cluster center of each cluster's feature vectors;
for each cluster, determining a plurality of cluster vectors from each cluster center to the cluster's feature vectors; and
for each cluster, determining an affine-linear space that includes the cluster's cluster center and that is spanned by a plurality of principal components of the cluster's cluster vectors,
wherein applying the classifier to the three or more basis test images is based on a distance between each test vector of the three or more basis test images and a nearest affine-linear space.
21. The inspection system of claim 13 , wherein forming the classifier comprises:
minimizing a norm of distances from a affine linear space to a plurality of feature vectors from the three or more basis training images;
discarding a set of the feature vectors for which a distance is highest; and
repeating the step for minimizing the norm of distance to refine a plurality of affine-linear spaces to better fit remaining ones of the feature vectors until the feature vectors are fitted to a final set of the affine-linear spaces within a predefined tolerance level.
22. The inspection system of claim 13 , wherein the classifier includes three or more dimensions of feature vectors that each has a transmitting image value, a reflecting image value, and a transformed image value based on transformation of a transmitting or reflecting image value.
23. The inspection system of claim 22 , wherein each transformed image value is determined by one of the following: ∥∇T∥, ∥∇R∥, ∇ 2 T, ∇ 2 R, wherein T is a transmitting image value and R is a reflecting image value.
24. The inspection system of claim 13 , wherein the one or more operating modes are selectable to differ in one or more of the following qualities or quantities: a reflecting mode, a transmitting mode, a brightfield mode, a darkfield mode, an illumination numerical aperture, an illumination pupil pattern, an illumination polarization, a collection numerical aperture, a collection pupil filter amplitude and phase, a collection analyzer, and a focus.