The mask inspection and defect classification is a crucial part of mask preparation technology and consumes a significant
amount of mask preparation time. As the patterns on a mask become smaller and more complex, the need for a highly
precise mask inspection system with high detection sensitivity becomes greater. However, due to the high sensitivity, in
addition to the detection of smaller defects on finer geometries, the inspection machine could report large number of
false defects. The total number of defects becomes significantly high and the manual classification of these defects,
where the operator should review each of the defects and classify them, may take huge amount of time. Apart from false
defects, many of the very small real defects may not print on the wafer and user needs to spend time on classifying them
as well. Also, sometimes, manual classification done by different operators may not be consistent. So, need for an
automatic, consistent and fast classification tool becomes more acute in more advanced nodes.
Automatic Defect Classification tool (NxADC) which is in advanced stage of development as part of NxDAT1, can
automatically classify defects accurately and consistently in very less amount of time, compared to a human operator.
Amongst the prospective defects as detected by the Mask Inspection System, NxADC identifies several types of false
defects such as false defects due to registration error, false defects due to problems with CCD, noise, etc. It is also able to
automatically classify real defects such as, pin-dot, pin-hole, clear extension, multiple-edges opaque, missing chrome,
We faced a large set of algorithmic challenges during the course of the development of our NxADC tool. These include
selecting the appropriate image alignment algorithm to detect registration errors (especially when there are sub-pixel
registration errors or misalignment in repetitive patterns such as line space), differentiating noise from very small real
defects, registering grey level defect images with layout data base, automatically finding out maximum critical
dimension (CD) variation for defective patterns (where patterns could have Manhattan as well as all angle edges), etc.
This paper discusses about many such key issues and suggests strategies to address some of them based upon our
experience while developing the NxADC and evaluating it on production mask defects.
The mask inspection and review process is a vital part of mask preparation technology and consumes a significant
amount of mask preparation time. As the patterns on a mask become smaller and more complex, the need for a
highly precise mask inspection system with a high detection sensitivity and low number of false defects becomes
greater. A low number of false defects is desirable as the results of the mask inspection are typically reviewed
manually by an operator in the mask shop. However, due to various reasons, the probable mask defects identified by
any mask inspection machine could include significant number of false defects. The false defects could be due to
registration or focus errors between the defect and reference images (Die-to-Die or D2D comparison), CCD
(Charge-coupled device) errors in the camera, noisy pixels etc. These false defects cannot be ignored and require the
operator to review them manually before classifying them as false defects. This takes valuable time and effort of the
mask inspector and increases the turn-around-time of mask inspection.
We propose a software tool which automatically detects most of the false defects generated due to registration and
CCD errors in the mask inspection system. It is quite common to find several thousands of defects (real as well as
false defects) during mask inspection. We have observed that significant percentage of these false defects are due to
registration and CCD errors in defect and reference images during D2D inspection. Automatic detection of
registration and CCD errors requires image processing to be done on the defect images. This process is typically,
time consuming. However, image processing algorithms are well suited for parallelization.
We explore the use of GPUs to speed up the false defect detection process by analyzing the defects in parallel on
multiple cores of a GPU. In addition, GPUs are inexpensive, readily available and can be plugged in to any desktop
computer which makes it easier to adopt. The proposed GPU based parallel false defect detection feature is
integrated into Mask Defect Analysis tool - NxDAT1.