The results of the multi-vari study proved to be insightful. For example,
the problem of low board-level capability was to a large extent primarily due
to within-board and board-to-board (within rack) variance differences.
However, the within-board and board-to-board averages were relatively
This piece of quantitative evidence gave the team considerable awareness
into why previous experimentation conducted months earlier failed to yield
substantiative answers. These failures resulted because board-level means were
used as a basis of analysis - the multi-vari study revealed that the problem
was constrained to the variance in certain areas of the PCB and plating tank.
Secondly, the experimental factors historically used for experimentation were
temporal in nature (time related) - the multi-vari study suggested that the
problem was positional in nature (focused within and between boards).
Consequently, the experiments were improperly vectored with respect to the
problem parameters and performance metrics. Not only did the multi-vari study
surface the major sources of variation, it provided a vital base for
brainstorming potentially influential process variables.
Figures B-21 and B-22 display the actual multi-vari charts developed by the
characterization team. Each datum within Figure B-21 and Figure B-22 consisted
of N = 120 individual plating-thickness measurements.
As seen in Figure B-21, the within-board and across-board averages were
relatively consistent. Figure B-22 reveals a strong, repeatable pattern in the subquadrant means
indicating the lower portion of each board received more copper on the average
than the upper portion, irrespective of rack position. It is also of note that
those boards with the largest plating averages such as boards 1 and 3 occupied
rack positions within the lower portion of the plating tank.
Given the multi-vari charts presented in Figures B-23 and B-24, it became
apparent to the team that copper plating uniformity was the major problem. A
careful analysis of Figure B-23 revealed that the problem was primarily con-
strained to the lower half of boards 1 and 3. Board number 2 also displayed
the same symptom to a lesser extent. Again, the problem appeared to be focused
in those rack positions within the lower portion of the plating
In terms of the test coupons, the averages were highly repeatable from
board-to- board, but not within board; however, several of the test coupons
were reasonably representative in terms of the standard deviation. Given that
coupon 6 was the buy-off vehicle, it was apparent (refer to Figures B-21,
B-22, B-23, and B-24) that the buy-off coupon was not representative of the
intelligent board area.
The team drew three major conclusions from the multi-vari exercise.
Greatest leverage could be gained by focusing on
the reduction of within- board and within-rack variation (positional
The primary performance metric for subsequent
investigations should be the response variance.
The lack of coupon-to-board agreement was primarily
a design issue thereby removing it from their immediate span of
In terms of the variance, it was obvious that a severe
problem existed (refer to Figure B-24). Not only was there a wide range of
standard deviations from board to board relational to the intelligent area,
but the problem appeared to be particularly acute within the low half of the
PCBs (quadrants 3 and 4). In referring to Figure B-24, it appeared that the
change was in the form of a sudden shift. It was also apparent that the
standard deviation was not highly repeatable across boards; however, this
concept did not hold true for all of the PCBs. Using Figure B-24 as a guide, it can be seen that the standard deviation was quite
consistent across the intelligent area of board number 4. This particular
observation created controversy within the team.
As before, the characterization team reasoned it was quite possible that
the undesirably low capability ratio was largely due to the wide range of
pooled through-hole measurements. When compiled into a single data set, the
board- level standard deviation would be consistently higher than the
individual through-hole standard deviations, therefore driving a low
Relating to the coupons, repeatability was moderately good across boards;
such repeatability was absent from coupon to coupon within a board. The
board-to- board stability of the buy-off coupon 6 was high; however, it was
not representative of the intelligent area standard deviation. Again, coupon 8
seemed to be the most representative.
Based on these observations, the team concluded that coupon number 6 was
generally not representative of the intelligent board area in terms of central
tendency or variability: therefore, the coupon-to-board correlation would
expectedly be low. The team further concluded that the given composite
intelligent area capability was controlled primarily by within-board
variations and not board-to-board variations. Furthermore, the problem of
non-conformance was more related to the variance than the mean.
As a consequence of these conclusions, the characterization team determined
that the variance should be the primary statistical response during variable
identification and optimization. It was decided that the optimization effort
should focus on the individual subquadrants within the intelligent board area.
In other words, the team determined that if the optimization study was to
"bear fruit," each subquadrant would have to be optimized, but in a collective
fashion. Finally, it was believed that if the key process variables were
related to each of the subquadrants, leverage could be realized in terms of
yield enhancement at the board level.