You are on page 1of 4

JFAPBC (2006) 3:35-38 © ASM International

DOI: 10.1361/154770206X107280 1547-7029 / $19.00

F•A•I •L•U•R•E P•U•Z •Z•L•E•S by David L. Burgess

Lessons from Sudoku

F
ailure analysis requires a work- • Each of the nine mini-grids must also be placed in the third column,
ing knowledge of physics, include each number 1 through 9. and no more than three squares are
technology, materials, manu- There are no more rules. possible sites. Note in the puzzle in
facturing practices, analytical tech- Fig. 1 that columns A and C each
niques, and more. It is difficult to One possible way of starting a contain an 8. Column B contains
imagine what failure analysis would Sudoku puzzle is to check each set of only a 4 and a 1. An 8 must be placed
be like if all the laws of physics were three columns to identify those that in column B. Where is an 8 possible
perfectly apparent and all the chal- contain a given number in two of the in column B? Not in the top left
lenges of sample preparation and three columns. That number must mini-grid; not in the middle left
testing were removed. Elimi- mini-grid. An 8 in column B is
nating those considerations, fail- possible only within the bottom
ure analysis would be more like mini-grid. B9 is occupied by the
Sudoku. Sudoku is a simple number 1. Row 7 already has an
number puzzle enjoyed by 8 at G7; so the only possible
children and adults. The sim- square for an 8 is B8. Similar
plicity of Sudoku compared to reasoning allows unique place-
failure analysis is the point. Al- ment of a 4 in A3. Repeatedly
though Sudoku is starkly simple, testing each number and each set
the process is challenging to the of columns and rows results in
best of us. Good and bad prob- filling in more squares. However,
lem-solving practices are clearly that tactic eventually stops
exposed by the experience of solv- producing results. The diagram in
ing a few Sudoku puzzles. Some Fig. 2 shows the puzzle pro-
insights that are brilliantly clear gressed to that situation. The
in Sudoku are equally true in fail- question “Where in row x is the
ure analysis. Fig. 1 number y possible?” is no longer
The Sudoku universe is a nine- effective. A new question, or at
by-nine array of squares. The least a new twist, is required.
array is divided into nine three-
by-three mini-grids, as shown in A search for fertile squares does
Fig. 1. At the start of a puzzle, reveal new possibilities. For
some of the squares contain a example, the middle right mini-
number between 1 and 9. The grid has five empty squares. The
object of the puzzle is to fill in unassigned numbers are 1, 2, 3,
the blank squares with numbers 4, and 7. Row 6 contains a 3 at
1 through 9. Labels are attached C6. Column G contains a 3 at
to the puzzle in Fig. 1 to identify G2. That eliminates row 6 and
columns and rows for discussion. column G as possible squares for
a 3. An 8 and a 5 occupy two of
There are only three rules.
the remaining squares, leaving
• Every row must include each only squares H4 and I5 as
number 1 through 9. possibilities for a 3 in the middle
• Every column must include right mini-grid. Is there any way
each number 1 through 9. Fig. 2 to reduce the possibilities to one?
This article is adapted from an article of the same name that originally appeared in Electronic Device Failure Analysis, ASM International, 2006, 8(1),
pp. 25-28.

Journal of Failure Analysis and Prevention Volume 6(3) June 2006 35


Lessons from Sudoku (continued)

Yes; consider row 5 on its own. Row Sudoku puzzle far enough to see problem could be anything and
5 must accept a 3 somewhere. Where, what it is all about. anywhere. Simply using available
if not I5? C5 is not possible, because information to state the problem as
column C contains a 3 at C6. (Be- What are the lessons for failure precisely as possible eliminates a
sides, that 3 satisfies the quota for analysts? Let us start at the beginning: few possibilities. In a process similar
threes in the middle left mini-grid.) • In Sudoku there’s no need to verify to Sudoku, extend initial data with
E5 is not possible because of the 3 in the initial data provided, as there is additional observations that sharpen
E8. Column G was previously in failure analysis. But beyond that, the problem definition. Some pro-
eliminated because of G2. If there is Sudoku puzzles and failure analyses gress will come quickly. Restate the
going to be a 3 in row 5, it must be start in the same way: Survey the problem using the new information
in column I. Column I, row 5 is a 3. information provided to get a feel to better define the “when,” “what,”
Note: It does not matter that other for the problem at hand. Initially, and “where” associated with the
numbers might fit in I5. If I5 is the in failure analysis, the cause of the failure.
only possibility for a 3 in row 5, • Continue to refine the problem
then I5 is necessarily a 3. No definition by contrasting the
more data is required. “when,” “what,” and “where”
Continuing still from the associated with “not the prob-
diagram in Fig. 2, look at the lem.” Surprises in the “not the
lower middle mini-grid. One of problem” category can change the
the numbers required in that direction of the analysis. The
mini-grid is a 4. All three squares universe will now fall into one of
in column F are filled and not three categories: “the problem,”
available. Row 9 is not possible “not the problem,” and “unclassi-
because of C9. Column D is not fied, could be or not, don’t know.”
possible because of D2. Therefore, The task of the failure analyst is
the 4 goes in the only remaining to bring the problem into clear
possibility, E7. focus by reducing the unknown
areas separating “problem is” from
Now, with E7 filled, the only “problem is not.” Reclassifying
square in the lower middle mini- things in the “don’t know” cate-
grid for a 9 is E9. Still using the Fig. 3 gory into “part of the problem”
diagram in Fig. 3, the only square or “not part of the problem” auto-
for a 6 in the lower left mini-grid matically sharpens the problem
is A9. That leads to placing a 6 definition. One step at a time, the
in B4. Therefore, a 6 must go in initial nebulous problem that
C1. That brings us to the diagram defied solution will be trans-
in Fig. 3. formed into a more defined prob-
Column C needs only a 1 and lem with a solid solution.
a 2 to be complete. The lower left • Depending on the problem,
mini-grid contains a 1, so the 1 physically isolating an anomaly,
in column C must be at C5. C5 characterizing a failure mode, or
is a 1. The only remaining square statistically correlating a cause
in column C—C7—must be a 2. with a failure mode may be the
Square D8 is also a 2. All the key to simplifying the problem.
entries to this point are shown in In every case, enjoy the problem
Fig. 4. and be attentive to all its details.
Enough about Sudoku. We Push back thoughts of the final
have followed through this one Fig. 4 conclusion. Just concentrate on

36 Volume 6(3) June 2006 Journal of Failure Analysis and Prevention


defining the problem from dif- ther. He may also question the into an observation that supports a
ferent points of view. interpretation of obser vations desired output. On the other hand,
• Just as in Sudoku, there will be already made. Progress is built do not discount an observation that
times when progress stalls and data upon all the facts developed during does not fit an otherwise satisfying
seems insufficient to proceed. The the analysis. A colleague (or you, for conclusion. Trust that all the facts
first impulse should be to pause, that matter) must believe each will fit and make sense when the
have a cup of tea, then take a fresh observation to commit full energy problem is understood. An analyst
to solving the problem from that with an above-average grasp of
foundation. You are physics and tools will very likely be
“Forcing the analysis in a predetermined asking for total trust in able to sell, in good faith, a faulty
direction will take longer than a more logical your discoveries to that conclusion. (Even the most gifted
step. Forcing the analysis to a conclusion before point. You should analysts can come to believe a false-
the problem is sufficiently defined is pointless.” therefore expect and hood.) Recognizing and correcting
welcome detailed a mistaken conclusion may take a
questions to support great deal of time and money.
look at the data. Enjoy the chal-
your characterization of the prob- • In Sudoku and failure analysis, for
lenge of finding valuable infor-
lem thus far. Simply reviewing the the reason above, it is wise neither
mation in the data. Failing that,
steps can be an enlightening to guess nor assume. Every piece of
look for an answerable question in
exercise. Good notes will be data is important and must be
the data—then do what must be
essential. Your failure analysis accurate. However, failure analysis
done to obtain the answer. In
colleague may believe what you say, is not a game. Statistics are not part
Sudoku, many questions are
but by nature, he trusts data more of Sudoku, but statistical character-
fruitless at any given stage, but the
than opinion. Where possible, istics are important, real, and useful
answer to the right question is there
provide direct evidence, not only a in failure analysis. Statistical aspects
for the picking. The data will
conclusion or interpretation. (This of a failure analysis problem should
dictate the direction of the analysis.
is especially important when not be overlooked.
Forcing the analysis in a predeter-
analysis is a shared effort among
mined direction will take longer • Because the Sudoku world is small
different labs and companies.)
than a more logical step. Forcing the and defined, it may be correct to
analysis to a conclusion before the • In Sudoku, a square filled in say, “This square cannot be a 4, so
problem is sufficiently defined is incorrectly will go unnoticed. The it must be an 8.” The real world is
pointless. Just as the puzzle solver process will continue nicely, fuzzier. It may be a slight exaggera-
concentrates on each next number, happily building on false infor- tion to say that you don’t know
one at a time, an analyst should look mation. More erroneous numbers what it isn’t until you know what it
for one more defining fact. That will be derived from the first. is. But be very careful. Being asked
fact will lead to another. The puzzle Erroneous numbers will be consis- to comment on a photo of an inte-
and the failure analysis problem tent with one another, creating a grated circuit with massive amounts
each will reveal its secrets in turn, false confidence that all is well. of melted metal and silicon, some
one possible step after the last. Even Only at the very end will it become analysts offered the comment, “I
if the boss and others are pressing obvious that a mistake was made. don’t know what it is, but it defi-
for a conclusion, the next answer- If a mistake is found, it may be nitely is not ESD.” (Electrostatic
able question is the quickest path necessary to start over, reverifying discharge, or ESD, was in fact the
to that conclusion. every number. Failure analysis must initiating cause.)
• A different pair of eyes may see comply with a more complex reality, • Even though Sudoku rules are
possibilities in the data that you do and one error may lead to erroneous extremely simple, puzzle-solving
not. A colleague looking from a conclusions. The mistake may not skill can be enhanced by studying
fresh point of view may move the be obvious, even at the very end. So a book such as Carol Vorderman’s
problem along by recognizing be careful with each step. Avoid How to Do Sudoku. This book, and
patterns that can be developed fur- reading unwarranted significance others like it, discusses how to look

Journal of Failure Analysis and Prevention Volume 6(3) June 2006 37


Lessons from Sudoku (continued)

at available data in different fully using the tools and data


ways. It discusses how to rec- already available. Even though
ognize patterns. Technical Sudoku is so simple, it is hum-
papers focused on the failure bling to realize that progress from
analysis process are few and may a seemingly impossible situation
not be identified as such. usually does not involve a new,
Check, for example, “Failure sophisticated way of reasoning.
Analysis of Steel Bearing More likely, the key is using
Plates” in the December 2005 information already available and
issue of this journal. This paper finding the right question. Both
does not discuss how to perform Sudoku and failure analysis are
any of the failure analysis tests. about the questions.
It does demonstrate critical
Just for the record, the com-
thinking about specifications
pleted puzzle is shown in Fig. 5.
and how to make conclusions
from test data. It shows how a Fig. 5 Free Sudoku puzzles are
meaningful conclusion can be available online at http://
made with less data than necessary “how much,” and their associated www.websudoku.com.
to determine root cause. The paper “nots.” The thought process for
is about how to get everything pos- problem solving in general and fail-
sible (but no more) from the data. ure analysis specifically is a learned
skill deserving recognition and David L. Burgess, Accelerated
• All the questions involved in
more attention. Analysis, PMB Q-310, 80 N.
solving a Sudoku puzzle are varia-
Cabrillo Highway, Half Moon
tions of “where” and “where not.” • New tools are important in failure Bay, CA 94019. Contact e-mail:
Defining a failure mechanism analysis. Do not let the lure of davidburgess@att.net.
includes “what,” “when,” “where,” magical new tools distract you from

Who do you want to reach today?


Advertising information on page 28.

38 Volume 6(3) June 2006 Journal of Failure Analysis and Prevention

You might also like