Page 50 - ChipScale_Sep-Oct_2020-digital
P. 50
Why acceleration
models?
I n r e li a b i li t y
tests, the most ideal
situation is to have
the test conditions
very close to the
us e (o p e ra t i ng)
c o nd i t i o n s. where F o , R o , h o , x o , θ o , and β o are the
However, because life distribution, reliability, failure
of time-to-market rate, time-to-failure, characteristic life,
and cost savings, and Weibull slope, respectively, at the
t h is is a l m os t operating conditions.
impossible. The Linear acceleration. When η = 1, i.e.,
practical way is linear acceleration [7]
to run accelerated x o = αx T
t e s ts wi t h then we have
increased intensity, β o = β T
a n d re al i s t i c θ o = αθ T
sample sizes and
test times. (Most
reliability tests,
t h e re f o re , a re
accelerated tests.)
In this case, the
Figure 8: Weibull plots of a 208-pin PQFP with 10% and 40% resistant change p r i c e t o p a y i s It can be seen that, for a linear
failure criteria.
to constr uct the acceleration (η = 1), the Weibull CDF
Figure 8 shows the Weibull plots of a acceleration factors to map (transfer) plots of the operating condition and
208-pin plastic quad flat pack (PQFP) the failure probability, reliability testing condition should have the same
with a lead-free solder paste subject to function, mean life, and failure rate Weibull slope (i.e., the shape parameter
- 40 125º C t her mal cycl i ng. A from a test condition to the service remains the same), and the characteristic
computer stored all the resistance operating (use) condition. Acceleration life is different by a factor of the
measurements. Let’s define one failure models are needed for determining the acceleration factor. It should be pointed
criterion as 10% resistance increase and acceleration factors [4,5,7]. out and emphasized that the same Weibull
the other as a 40% increase. It can be As mentioned earlier, the best-fit slope is not an assumption but rather just
seen from the Weibull plots that: 1) the Weibull life distribution of a set of lead- came out naturally for linear acceleration
life distribution is many times different free interconnects is [7]. So, if two or more different test cells
from the failure criterion based on the (with different temperature conditions)
10% resistance increase and the 40% of the same lead-free interconnects
resistance increase; 2) for the same yield very different Weibull slopes, then
percent failures, the life taken from the either the linear acceleration model is
failure criterion based on higher whe r e F T , x T , θ T , a nd β T a r e t he not good (i.e., η ≠ 1), or the Weibull
resistance increases is longer; and 3) life distribution, time-to-failure, distribution is not fitted with the test data,
as expected, there are more failures with characteristic life, and shape parameter or both. Usually, the Weibull distribution
the failure criterion based on lower (or Weibull slope), respectively, under adequately represents fatigue failures,
resistance increases. the test conditions. Let’s consider the therefore, nonlinear acceleration models
Another important factor affecting following simple acceleration model may need to be considered. All the papers
the results of reliability testing is data (transformation or mapping) [7]: in the literature have been using the
extraction. During tests, how many linear acceleration, i.e., η = 1, or x o =αx T
measurements should we take at each or N o =αN T , where N o = x o and N T = x T .
cycle? It could be four measurements where x o is the time-to-failure at In most solder joint reliability tests of
or eight measurements. In order to operating condition, α is the acceleration integrated circuit (IC) packages on PCBs,
avoid false failures, it is recommended factor, and η is a real number larger than the luxury of running multiple thermal
to compare the measurements of every zero, e.g., 8.888. Then, we have cycling conditions is not available (due
channel at every two sequential cycles (no to the test time, chamber occupation,
matter if it is four measurements or eight and manpower). Therefore, acceleration
measurements). A three-cycle moving models are required to predict the life
average method is recommended. Please distribution and failure rate of the
read [5] for more details. solder joints under anticipated service
conditions. In the literature, for SnPb and
48
48 Chip Scale Review September • October • 2020 [ChipScaleReview.com]