=================================================================================
High
level of diagnosis noise, which could impact analysis
result. Root Cause De-convolution (RCD) is a statistical enhancement technology, based on a set of
assumptions, which reduces diagnosis noise by statistically evaluating the
existence probability of each defect suspect based on design
statistics [2]. Note that such an assumption used in RCD
may not always be valid. This RCD method provides an easy-to-understand defect pareto,
together with targeted Physical Failure Analysis (PFA)
candidates. Unfortunately, even the RCD analysis also has some assumptions and limitations, and its result cannot always be interpreted literally.
It's very common for scan diagnosis to contain
multiple symptoms and suspects. To remove the noise from the diagnosis, RCD is needed (Figure 4221a). Similar to a noise filter, RCD eliminates this noise and identifies the underlying
root cause distribution. The RCD algorithm with RCD pareto can:
i) Ranks all
suspected root causes so that we can prioritize all root
causes. We need to increase the PFA hit rate by ignoring
irrelevant locations with non-target root causes.
ii) Generates a defect distribution by calculating the probability of observing diagnosis results based on design-and test-weighted statistics. [1] Given a specific root cause, the probability of a specific suspect can be calculated by critical area per net segment per layer. For
instance, the probability of observing a specific suspect
will be equal to the suspect's critical area for the specific
root cause per layer, divided by the total critical area for all
possible suspects. The probability of seeing all the
suspects in the reports and can be calculated to determine
an overall probability number. By leveraging the suspect with the highest ranking
to the likely root cause, the noise-like
suspects can be easily eliminated.
iii) Automatically determines the underlying root causes in
represented devices from a population of failing test data
alone.
iv) Find the dies that are most likely to represent a root
cause of interest and sort the suspects in order of
probability for a particular root cause.
v) Provides more flexibility to handle the
dies with multiple symptoms and suspects.
vi) Allows to
select PFA candidate based on probability of specific root
cause.
Figure 4221a. Root Cause De-convolution (RCD).
RCD Analysis procedure can be:
i) Test the chips and collect the failure
information.
ii) Build a pareto.
The RCD analysis distributes the detected defects for the
whole population into a few pre-defined categories of yield
loss root causes.
iii) Build a contingency table.
Assuming random distribution of the defects, across
different sub-population (zones), RCD estimates the
expected and actual defects for each failure mechanism in
these zones. A difference between the expected and actual
values may indicate an excursion or zonal modulation for
the specific failure mechanism.
iv) Diagnosis simulation can
be executed to gather a list of possible failure locations within the individual failing devices.
v) Volume diagnosis can be
applied on top of a collection of diagnosis reports to generate a statistical summary of the different failures found within the individual devices.
vi) Focus on the major failure issues and pick out PFA candidates to understand the failure mechanism.
vii) Build RCD Suspects.
The RCD algorithm assigns a confidence level for
each suspect to be the real defect for each failure mechanism. This is especially useful for PFA selection
when the goal is to target on a specific failure mechanism.
Challenges for RCD are:
i) One main challenge of this RCD analysis flow lies in the
inherent ambiguity in the diagnosis results. [3] Due to the
inevitable structural and functional equivalence of faults in the
design, a typical diagnosis data often contains too many
suspects for each defect and it is not clear which one is the real
root cause. This in turn makes the volume diagnosis very difficult. Traditionally, one can count the total
number of times that a cell failure appears as one of the
suspects. However, the ambiguity can be
dramatically reduced and a clear failure scenario is painted by a machine learning algorithm based on
Bayes net model [2].
ii) Layout pattern root cause. This refers to a specific hard-to-manufacture layout structure
that becomes prone to fail at advanced technology node. Such prone to fail features occur more and more frequently as
feature size of the layout shrinks dramatically.
How to identify such layout features is an open question. Furthermore, the extremely large
number of potential layout patterns to be considered for statistical analysis poses a
challenge for statistical analysis.
iii) Cell internal root cause. This refers to cause of defect inside a standard library cell. At
more advanced technology node, library cells require more process steps and more complicated structures. The complicated structure of library cell and current limited
domain knowledge of the defect behaviors lead to a large number of
manufacturing defects and systematic root cause inside library cells, which have more
subtle defect behaviors.
============================================
[1] White paper: Root Cause Deconvolution - The Next
Step In Diagnosis Resolution Improvement, Geir
Eide, Mentor Graphics.
[2] Benware, Schuermyer, Sharma, Herrmann, “Determining
a Failure Root Cause Distribution From a Population of
Layout-Aware Scan Diagnosis Results”, IEEE Design &
Test of Computers, Jan/Feb 2012.
[3] Yan Pan, Atul Chittora, Kannan Sekar, Goh Szu Huat, You Guo Feng, Avinash Viswanatha and Jeffrey Lam, Leveraging Root Cause Deconvolution Analysis for Logic Yield Ramping, ISTFA, DOI:10.31399/asm.cp.istfa2013p0602, istfa2013p0602, pp. 602-607, 2013.
|