COMPREHENSIVE MODEL FOR HUMIDITY TESTING CORRELATION

D. Stewart Peck
-
3646 Highland St., Allentown, PA 18104
(215) 398-1222

ABSTRACT provides confident extrapolation to low stress con-

ditions, but with little expectation that anyone
This paper reviews all published life in humidity facility could mount a large enough effort to provide
conditions vs life at 85C/85%RH for epoxy packages; 61 data over the necessarily large range of stress, with
data points are used. An acceleration formula is known factors for repeatability.
described which provides direct extrapolation from test
results in autoclave tests at up to at least l40C to Another modeling and testing objective is to re-
l~-humidity down to below 30%RH. This formula compares duce the testing time from the usual 1000 hours at
favorably with previously-published formulae. The 85/85, particularly since median lives of present
possible effect of the Tg of the epoxy is recognized, epoxies are around 5000 hours, with very low standard
and other high-temperature and low-humidity effects are deviation, so that the 1000 hours tells little about
disclosed. Recommendations are made for future tests, the long-life properties of the epoxy, and can only
replacing 1000 hours at 85C/85%RH with a 20-hour test measure the freak level. Testing at a more severe
at 140C/IOO%RH, to reduce test time and increase test stress is becoming more common, particularly with the
usefulness. use of an autoclave above 100C and with controllable
percent saturation. Other techniques to shorten test
INTRODUCTION time, such as soaking before biasing or the use of a
prior salt exposure, have been reported, but these com<
Reliability testing in humid environments will plicate the extrapolation of the pure temperature-humi-
apparently be with us for a very long time. dity bias results to low-stress conditions, and these
Improvements in processing quality have essentially conditions are not considered here. This does not say
eliminated sodium contamination from complex chips. that they do not have advantages for the objective of
Better understanding of electromigration has provided reducing test time for a chosen test.
design limits which should essentially exclude problems
with this failure mechanism. Other mechanisms are It is the purpose of this paper to review all the
being controlled through continuing studies of published data on humidity testing over the entire
reliability physics, but new concerns may still arise stress range, both above and below 85/85 stress, in the
from failure mechanisms intensified by new VLSI design hope that either the existence or non-existence of a
rules. good correlation between autoclave testing and 85/85
could be established, and that at least a model could
And yet, the pressures of increasing pin-outs and be fitted to all the data from stress less than that of
size, together with the economics proven long ago, 85/85.
require that we continue to use plastic packages of one
form or another. The concomitant risk of failures The use of data from a wide range of stress
resulting from the effect of humidity on the chip minimizes the impact of sampling and testing variations
metalization will continue to require testing in high at anyone point. If autoclave testing provides valid
humidity and temperature for accelerated-stress acceleration, test times can be drastically shortened,
testing, to (1) evaluate package and chip materials and the frequency of testing can be improved, and predic-
design, and (2) provide efficient acceptance testing tions to low humidity conditions could be made, if at
on completed product. all, then from a more solid base.

The standard test for the effect of humidity has EVALUATION PROGRAM AND RESULTS
been, for many years, the electrically-biased test
of components in 85c temperature and 85% relative A survey was made of all published median-life
humidity. At the initial use of this test, it was data on electrolytic corrosion failures of aluminum
considered a design-and processing~control type of metalization (with the inclusion of one batch of data
test. Specification writers and device manufacturers on nichrome fuse electrolysis) in epoxy packages, from
felt that, if they could meet the usual requirement of available sources[l-14J, avoiding that data which did
10% failures in 1000 hours for this test, the devices not provide median life measurement at 85/85. A few
would operate as reliably as others in general-use exceptions were made where the 85/85 median could be
conditions. Attempts were not made to determine a estimated from nearby conditions with a minimum of
relationship between 85C/85%RH (85/85) test results and probable error. The 85/85 life being the base, the
expectation of life at some other environmental ratio Ro of observed median at some other condition to
condition. that at 85/85 was determined. As lives both longer and
shorter than at 85/85 were involved, this ratio
Since about 1970 we have seen the results of provides a consistent set of numbers which do not con-
various laboratories' efforts to understand such a flict. Ro is the Acceleration Factor (AF) from 85/85
relationship for plastic-encapsulated semiconductor to lower stress condition, but l/Ro would be the AF
devices. These efforts are frequently based on a small from higher-stress to 85/85; to show AF for both
number of tests at only a few conditions of humidity conditions on one plot would be confusing.
and temperature, which is not surprising considering
the long testing time for any condition much less Acceleration and Deceleration
severe than 85/85. A mathematical model is then
postulated which seems to fit these points. Serious The IEEE Transactions on Reliability, in Nov 1970,
thought should tell us, of course, that there must be published definitions which include:
variation around each of the measured median lives, in- "Acceleration Factor: The ratio between the times
cluding sampling variation, changes in the encap- necessary to obtain a stated proportion of failures for
sulation properties from time to time, contamination, two different sets of stress conditions involving the
and many other factors. Hence, a model which perfectly same failure modes and/or lIlechanisms."
fits a few data points may still not be correct. So
there is a continuing effort to find a model which Note that this definition does not distinguish between
CH2256-6/86/0000-0044$OI.00 c 1986 IEEE!IRPS 44
(t./t2) and (t2/t.), but calls either one an 100 .,.-------.------,----..----"""7\
-
acceleration factor. In this author's opinion, this is I
a:
a good practice. If, in going from an observed life at
low stress to a shorter life at high stress by a
factor, e.g., of 175, then 175 is the acceleration u 10 l = - - - - - 4 - - - - - - l - - - - - . A r r - - - - - . j
factor(AF). One divides the long life by 175 to obtain
the life at accelerated stress. If one then has an
.."'
~
observed life at the high stress condition and wants to ~

g~
predict life at low stress, it is undesirable to have
1.0
to remember that the Deceleration Factor (DF) is
0.0057142; it is only required to remember that the AF OJ
>
is 175, and one multiplies by it rather than dividing ~
oJ
by the OF. Only a single AF between two conditions is OJ
a:
required if one uses it appropriately. w
u.
0.1
:::;
Z
Also defined is: c(

"Accelerated Test: A test in which the appljed is

w
stress level is chosen to exceed that stated in the ::Ii

reference conditions in order to shorten the time fil .01

>
a:
required to observe the stress response of the w

o'"
item ... (identifying a common failure mechanism) ... " '" I
DECELERATION FACTOR• •ACCELERATION FACTOR

This implies that the life at the higher stress is the .00 1 '-:-...L.J...J.J...U.':'"--.J.---l...J...U.l.ill_..l.-LLW-L"':----L...J-JL.llWJ
.01 0.1 1.0 10 100
accelerated life, and is normally obtained by dividing CALCULATED MEDIAN LIFE
the long life by the AF. In the present review, the RELATIVE TO THI\T 1'.1" 85 C I 85% RH
ratio of median life at some other condition to that at
85/855 is used, giving an AF when treating lower
stresses, and a DF when treating higher stresses. The Figure 1 Ratios of life in reported humidity tests to
total AF between a stress below 85/85 and that above that in 85C/85%RH are shown, comparing the
85/85 would be the AF below 85/85 divided by the DF observed ratio Ro to the calculated Rc from
from the higher stress. the ~odel. The ratios for various test con-
ditions are shown for perspective.
Modeling and Model Results
Linearity of the regression was gauged by eye from
The modeling involved in this progr~ is to pos- Figure 1; distributions of HO/Rc are shown later.
tulate a relationship between life-and-temperature and Linear regression analysis provides the intercept on
life-and-humidity so that the product of the two the Ro axis at 0.917 + or - 0.064, and the slope as
separable factors between a given condition and 85/85 0.994 + or - 0.04, the tolerances being determined from
provides a calculated ratio Rc, according to the model, the standard deviation of error. Removal of one
which best .atches the observed ratio, Ro. A plot of maverick point changes these numbers to an intercept of
Ho va Rc is used to determine regression parameters and 0.982 and a slope of 1.026, with the s~e tolerances as
the linearity and uniformity of the regression. before.

The lognormal distribution of life was assumed, as There was also some SusplClon that the earliest
it was in all the data reviewed. The standard data may not have represented the best-controlled test
deviation estimated in all tests ranged from 0.4 to conditions, and that later data might give better
0.5, showing remarkable uniformity. Ho and Hc also results. For data published in 1979 or later, the
have lognormal characteristics, as does the ratio combination of E.=0.81eV and n=-2.66 has a correlation
Ro/Rc. The use of the life ratio allows the consistent coefficient of 0.988, with intercept and slope of 1.013
comparison of data from 5-volt and 70-vo1t operation, and 1.012, and with the best standard deviations of
from test structures and from commercial devices, and distribution above and below 85/85 (0.175 and 0.281
from different time periods. respectively). Optimization was not explored, for the
sample size for Rc>l.O is then only seven and more data
Figure 1 shows the plot of 61 data points, with Ro should be obtained before this is studied further.
on the vertical axis and Rc on the horizontal, the 45° There is the potential for reducing the confidence
line representing a 1:1 correlation. The relationship intervals for extrapolation to low-humidity conditions.
shown here is:
It is apparent, however, even from existing data,
time-to-failure {RH)n exp (Eajkt) that a single relationship can be used to relate
stresses both above and below 85/85, making viable a
where the exponent n in Figure 1 is -2.66, and Ea direct extrapolation from autoclave test results to
O.7geV. The resulting acceleration calculation is life expectations at low-stress conditions. Further,
relatively insensitive in the range of n from -2.5 to publication of more data at autoclave stresses should
-3.0 and of Ea from 0.77 to 0.8leV, using the be strongly encouraged, in order to define the model
regression parameters and the distributions of RotRc more precisely for modern test facilities, and to
above and below 85/85 (_edians and estimates of determine a measure of repeatability. The lines at the
standard deviation) as criteria. Thus there are seven lower left of Figure 1 show the OF obtainable at some
measures of fit, and not all are optimized at anyone COnditions. The acceleration of 130/85 over 85/85 is
choice of n and Ea. The following combinations are almost 20:1, and 100%RH increases that. It is
suggested as nearly equivalent (and optimuu) in interesting to note that, for all the fears about
results: humidity testing at or near lOOe or 100%RH, the data
show only marginal influence of these limits; 16 of the
Ea n correlation coefficient 61 data points are within 5% of these limits, and all
are within the range of the other data, with the
0.79 -2.66 0.986 exception of one point at 100%RH, from 1972, at the
0.81 -2.50 0.985 lowest relative value of the distribution in Figure 2
0.77 -3.00 0.987 below.
45
Uniformity of Model Match resulting distribution has an improved median of 0.956
and an s of 0.281. Deletion of these points does not
-
An even division of Ro above and below Hc reflects change the correlation coefficient of the model, but
a uniform match of the data with the model through the improves the Ro intercept from 0.982 to 0.999, and the
entire range of the data; a visual review of Figure 1 slope from 1.026 to 1.020. This suggests that, while
shows this. Figures 2 and 3 show the distributions of 100%RH is useful for reducing the time of routine
Ro/Rc for Rc<l.O and Rc>l.O respectively. The medians tests, it may be desirable to avoid this condition if
of these distributions are at 0.919 and 0.963, both one wants to optimize extrapolation to low stress
being satisfactorily close to 1.0 A further breakdown levels. Further, testing in modern equipments should
to smaller segments of the range of data would show be aimed at determining if this fine distinction is
significant at the present time.

10 High-Temperature Limitations
Rol Rc FOR
R c <1.0 Several factors may be involved in determining a
temperature limitation for testing. One of them can be
n= 42
disposed of quickly, for it may not be a real problem;
med=0.92
s= 0.295
investigators should be aware of it, in case it shows
up. An anomalous result appeared in testing of
"., silicone-molded devices with Ti-Pd-Au metalization at
"
0::
~ temperatures in the range of 150-2500 and at 80%
o
0:: 1.0
,/ saturation [15).
0"
~
0::
.
./ "" 400
/' 350

300

~ 250
<t
(.)
fJ>
... 200
{ ([
~~
.10
10 30 50 10 90 98 99.8
0.2
CUMULATIVE PERCENTAGE
'<w"t 150
z

~ I"'----
::::i """..:::
u
Figure 2 The distribution of deviations (Ro/Rc) from
the model for stress greater than (life ::! 100 .......
::>
shorter than) at 85/85.

10
!;;:
'"
W
11.
::.
w
.... 50
" ""'

RO/R c FOR t------

R c >1.0
n =19 10' 10' 10 3 10' 10' 10'
mad: 0.963 MEDIAN LIFE (HOURSl
s= 0.330
./

"
0:: ~ Figure 4 Examples of possible anomalous life be-
o / ~ havior at high temperature in 80% sat-
a: 1.0 uration (as seen in silicone encapsu-
0"
~ lations) .
0:: ,.
This is shown in three examples in Figure 4 as a
• reversal in the temperature effect, giving better life
at 2500 than at 2000. This may not be a problem in
present testing because of the difference in tempera-
ture capabilities of the materials. If, however,
improvements allow testing beyond 1500, this effect
.10
0.2 10 30 50 10 90 98 99.8
might possibly appear. It was not investigated at the
CUMULATIVE PERCENTAGE
time (1969), so there is no background for judging its
possible presence in other materials.

Figure 3 The distribution of deviations (Ho/Hc) from There are other evidences of notable behavior at
the model for stress less than (life great- 1500. Figure 5 shows an addition to figure 1, adding
er than) at 85185. data for 1500 results for several epoxy devices[15]. At
80% saturation these results, together with those at
distributions of similar uniformity. The data from 200e on silicone encapsulations, show both quite longer
stresses higher than 85/85 have a better distribution and quite shorter life than would be predicted by the
and a smaller estimate, s, of the standard deviation. present model which appears so effective up to that
The confidence intervals (90%) for the standard stress. These data were obtained in the late 1960's
deviation above 85/85 are 0.25-0.36, and are 0.26-0.46 and may simply reflect the lack of understanding at
for lower stress. These are not greatly different, but that time of proper testing procedures in autoclaves.
suggest that the distribution at higher temperature may One might guess, however, that anything done "wrong"
be somewhat better. This is not surprising, might tend to shorten life rather than lengthen it.
considering the shorter test times, with less Therefore, the regime of testing at 1500 and above, for
opportunity for test condition perturbation. epoxies, is one which should be re-examined when it can
In the data of figure 2, six results represent be done safely with respect to glass transition
lOO%RH conditions; if these results are eliminated, the temperature.
46
 ----------------~

about corrosion" , but data are not available. Some

models are offered to apply "down to 25"", but without
:r explanation. It is, of course, very difficult to take
a:
'$
100 the time to run a test at 10% or 20"RH at normal use
'" o-T<lOoC temperatures, or even at B5C. The lowest humidity
'" l(-T.;llOOC level in the data used here for epoxies was 50%, and
()
I I the 50-60% data are spread reasonably about 1.0 for
'"'" 10 L_
t- Ro/Rc.

r-
~

1 I '0
• 0°0 D
~
...:r The D.E.T. curves as shown by Koelmans[17) show
r l50Cl80'll. /0 0
that the drastic reduction in surface conductivity (vs
......> ~---<>-
0 1.0
- ; - ------l-
zoo,eo ., I f /f. ~ I RIl) takes place at about 5%; at this condition, at 150C
the life should be 36 times that at 85/85--an
~
., .,
" ~ ~t:XX0!'4-100C/8$% untestable time. Even at 20OC, the AF at 5% is 3.66,
W
II: 0.1
ZOO/80 U- --J..-. still a poor test time. At 20", however, the ratio to
w
u.
::::; -]~- ...~ -- llo,Cf8""'--- life at 85/85 is 0.91 at lSOC, 50 there is some
.
z
ffi .01 --
.4
i%

-~-~--
wl( If ,...- 120Cl'~%

fc--I~OCI e5~. ~'5Qr\8

1
\
opportunity to observe the actual life ratio if we can
otherwise be satisfied with the testing capability at
150C. At 150/10, the ratio should be 1.96, so that any
:i:
~ 200/X' I 2:00il' I
extension of life beyond that calculation would require
"w
>
o:ue
(j,il~/9Q
(ll I '
l
an unacceptably long test time.
ffi ~~OO/IOO DECELERATION FACTN( • • ACCELLRAT10N rACTOR

l:l .001 I-J....J...w...w..-..l....J....lJ..U.;.J-...LW-L;.w--l....l..JUJ..lill---i-LU.llill

o .001 .01 0.1 1.0 1Q() On the other hand, the question can be raised as
CALCU~ATED MEDIAN LIFE
RELATIVE TO THAT AT 85 C I 85 % RH
to the practical value of such investigation. In an
application at 30/20, the acceleration to 85/85 is over
4900; a SODa-hour epoxy median life in 85/85 becomes
2.5x107 in 30/20, and the failure rate will be less
figure 5 Comparison of Ro and Rc, including data at than 1 FIT for 300 years. At 40/20, the AF is 1877,
l50C and 200C showing variable behavior at providing a failure rate below 1 FIT for 50 years.
high temperature, and divergent results at These calculations do not even provide for any increase
low humidity. (ca 1969) beyond the calculated value, and already the question
is of only academic interest.
One effect of the glass transition temperature,
Tg, is indicated in Figure 5; two points, marked by APPLICATIONS
squares, represent data taken on product with a stated
Tg of 145C[6]. The 130C data at point (0.03, 0.07) Extrapolations from Autoclave Test Results
(Ro,Rc) is on the low side of the distribution, but not
outside it. The 150C data, shown at point (0.004, The consistency of the T and RH effects from T<85C
0.025) is almost a factor of ten below the calculation. through T>85C (and, perhaps more importantly, above and
The work of Vanderkooi and Riddell [16) in 1976 showed below 100C) suggests that there is no reason we should
that the permeability of epoxy by water vapor, while it not be able to extrapolate directly from results above
increases slowly as temperature approaches Tt, takes a 85C and predict failure rates at the low humidity
30~ jump at Tg, and then proceeds faster than before stress of an application (keeping in mind the pessimis-
as a function of temperature. This would lead one to tic calculation below 25 or 30%RH). If the model pro-
expect a more rapid access of Doisture to the aluminum posed here holds up with further testing, or can be
at and above the Tg, resulting in faster corrosion and fine-tuned satisfactorily, then the distribution of
shorter life. Figure 2 can be translated, according to the model,dir-
ectly to the low-stress condition. It might also be
The paper of Merrett et al[7], however, shows expected that, with improved equipments and operating
results on the product of Mfr. C at 131/90 (as practices, and known Tg's, even a tighter distribution
estimated from the other data) as falling well within wi 11 be found.
the range of the model, although with a stated Tg of
127C; also, the presumed humiditY"reduced Tg's for The upper 90% confidence level of the standard
Mfrs. B and f (128C and 125C) did not affect the deviation of results about the calculated ratio He, for
results at 131/90 enough to put them outside the expec- stress above 85/85 is 0.360 (Fig.2). For this
tation from the model. In the latter cases there is lognormal distribution, then, the central 90% of the
the possibility, indicated by people in the epoxy distribution is included within a range around the
manufacturing business, that the supposed reduction of median defined by a multiplier and divisor of 1.8 (1.7
the Tg in huuidity exposure is not a real effect with with deletion of 100%RH data). The measured life in
respect to .oisture permeability, and that the original this condition can be extrapolated by the model to a
136C measurement still applies; then, the 131C tests low-stress condition, and the median life there should
might not be expected to have an effect. There appears be within a factor of 1.8 from the calculated value,
to be some uncertainty about the effect of Tg, or the with 90% confidence. The regression line passes
teaperature interval of the effect, on the corrosion through Ro.axis at 0.982, and the central value of the
test results. Tests should be made well on either side distribution of Ro/Rc is at about the same point, say-
of the measured Tg prior to hUJIidity exposure, to ing that the regression bypasses 85/85 by only 2-3%,
establish the degree to which this epoxy parameter (or and matches the low-stress extrapolations of Figure 1.
so~ modification of it) will liDit testing tempera-
ture, or invalidate the model. Such predictions can not be applied to low failure
percentiles, which .ay include freaks which can vary in
Low-HUJlidity Effect a given sample from the freak level of the population,
and has been seen to be as high as 10% in some
Something else can possibly be learned from test- published data. Similarly, estimates of median lives
ing at high temperatures --- an answer to the question: from test data only to early failure percentiles can be
"at what low %RIl does corrosion effect diminish more made only after careful treatment to eliminate freaks;
rapidly than the mode11" The statement is often made otherwise the median estimate could be grossly
that "below some level of %RH we don't have to worry distorted.
47
 Since the data used herein were obtained by
different groups of experimenters over a period of over
The answer also depends on what the user of the
product needs[19]. If the application is in a severe
-
10 years, it is amazing that even this degree of con- environment of 40C/95%, the 4000-hour (85/85) product
sistency could be found. The 90% confidence for median will have a failure rate of 500 FITs in seven years,
life about the sample value, from a homogeneous and 8000 FITs in 10 years. A product similar to the
population, for samples of 50 (which most of these above, but with a different epoxy date, was tested in
were), and with a of 0.5, is a factor of 1.12, leav- the same time period, the 121/100 test showing a
ing a factor of 1.6-1.7 to account for the accumulated probable 85/85 median life of 7000 hours. This would
vagaries of these humidity tests. And this does not have operated comfortably in the 40/95 environment, and
even consider the variability in the 85/85 test which with a tested performance in only 200 hours in the
is the base for the calculations. There a temporary 121/100 conditions. An appropriate test condition for
drop in temperature of 4C could result in condensation that application should be able to distinguish between
which could reduce life in a chamber not designed to the two products.
protect against condensation on the devices.
There are now some data available on epoxy and die
There is an unquestioned usefulness for the model surfaces which will provide from 10,000 to over
shown for direct extrapolation from autoclave results 20,000 hours median life in 85/85. These times put
to low humidity stress, for product similar to that 85/85 testing out of the question for control or
included in this review. (Some data from phenolic and acceptance purposes. At 140/100, however, the 20,000--
silicone encapsulations also match the model, but were hour equivalent is only 400 hours, and a test allowing
not included in the calculations). Significantly less than 50% defective could be established in a
different insulating materials and encapsulations shorter time.
should be evaluated separately.
If Military conditions are a real objective, then
A paper by O. Hallberg[18) came to light as this perhaps something like 38C/98% might be an extreme
was being finished. It used 51 points of data from condition, in lieu of a specification. For a maximum
different sources, in a similar fashion to that of 50-FIT failure rate for 10 years in this environment, a
Figure 1 here, but without that statistical comparison median life can be determined [19), and the
of the models discussed. The comparison being acceleration factor indicates that the corresponding
suggested, Appendix A is added to treat this subject. median life at 85/85 would be 24,000 hours. This could
be established by a requirement of 10% failures in 465
hours at 135/95, or 350 hours in 140/95 (or 60 hours in
10,000 150/100 with 5% defectives allowed).

With the indication in these data of the possible

>:: need to stay under the Tg of the epoxy, these
0:
,-
(192)
"" - ~-
f---
conditions may put pressure on the selection of
materials for the encapsulation. The tremendous ad-
""
,xh44l' vantage, however, of reducing test times from unreach-
'" able numbers to quite modest numbers of hours, might
'" (96)1"" warrant some cost for the Tg. Even if the material
!a:
~ 1,000
II systems are not yet available to meet the most severe
::>
o
>::
~
z
~X/(48)
== • RESULTS AT 85 C/85%
x RESULTS AT 121 C/100%,
requirements, the testing facility should be validated
and calibrated, since it seems within the industry cap-
ability.
~ WITH ( ) HOURS AT
~
5
I 121 C / 1000/0
According to the present model, acceleration fac-
aw
tors over 85/85 results include the following:

Condition Acceleration
100
0.1 0.5 2 10 30 50 70 90 98 99.599.9
121/100 16
CUMULATIVE PERCENT DEFECTIVE
135/94 30
140/94 40
140/100 50
Figure 6 Composite life distribution from 85/85 150/100 77
and 121/100 tests.
The equivalent of 1000 hours at 85/85 could be 20
Establishing Humidity Test Conditions hours at 140/100, allowing this test to be completed
within one day. This would only be a measurement of
Figure 6 is an exampl~ of the combined test the freak level in the product, but could very well be
results of both 85/85 and 121/100 conditions (on the a specification requirement, since it will control the
basis of the equivalent time at the 85/85 conditions) early failure rates. In a relatively short time (5 to
from many samples from the same production period of a 20 days) one could describe the median life of any
semiconductor product. During the 1000 hours at 85/85, present epoxy system, or of one which might meet
only the freak level of the product was measured (but Military requirements.
the same result could have been obtained in 48 hours at
121/100, or in 28 hours at 135/100). Further, the CONCLUSIONS
important information that the product had a median
life of about 4000 hours at 85/85 was completely First, there is available here a relationship
unavailable except by testing up to about 3000 hours at between temperature, humidity and life for electrolytic
85/85, 200 hours at 121/100, or 84 hours at 135/100. corrosion of aluminum metalization, which can be used
to extrapolate autoclave testing results, accounting
Which test is the more desirable? The shortest for the possible effect of the epoxy Tg, directly to
test provides the quickest feedback for control, the expected life at low stress down to about 30%RH, below
greatest throughput of samples, and the greatest incen- which life may be longer than indicated by the model.
tive for frequent testing. Avoiding testing at 100%RH may give a slight
48
 ------------~_=s!

improvement in the confidence limits of the extrapola- lation to the low-humidity-atress region.
tion, from past data; the possible effect with modern
equipMents should be explored further. The present model provides the tightest distri-
bution of fit to all the data used.
Second, this relationship allows the establishment
of very-short-time tests to replace the present 1000- 10
hour, 85/85 tests, and this should be done i~ediate~
More test results should be reported at both below and
above 85/85 stress. Low-stress tests are needed to 'l
improve the statistics in that area, and the high- r- Re >1.0
~"3
stress tests should emphasize measures of repeat-
ability, particularly at 100%RH. There appears to be ~
~yl
no other reason not to use this condition for routine
V~
testing, except for the reproducibility of ratio to u
ex:
V
85/85, compared to that of lower RH. ;p 1.0
o
Third, the existence of epoxy packaging systems ~
for semiconductor devices with median lives greater a: ./ -'
than 20,000 hours has been suggested. Such devices '/
could have acceptable performance in Military applica- W
tions, and the autoclave results discussed here /'
indicate that they could be qualified in reasonable
tiDe periods. Formulation of humidity- test require-
ments for these applications may now be timely. .1
0.2 10 30 50 70 90 98 99.8
APPENDIX A CUMULATIVE PERCENT

A paper by O. Hallberg[18] also reviewed publish- Figure 8 Distributions of deviations from the model
ed data on the ratios of median life in other for the models of Figure 7, for stresses
conditions to median life in 85/85, and plotted 51 less than 85/85.
points of data according to the (RH)2 model of
Lawson(20,21] and the (T+RH) model of Reich and 10
Hakim(22]. This work was effective in reducing the
number of models, but did not use distributions of the
observed ratios about the calculated ratios as a I--- t- R c < 1.0
measure of fit, which is felt here to be the better "3
method of cOlllparison, together with the regression par'- Vl~
ooneters.
l:?"2
~~
a:u
REGRESSION PARAMETERS o ~ ./
log Ro = a + b (log Rc)
a;: 1.0
0"
~ '7
CORRELATION a b
II:
,......-........ l./V
".
COEFFICIENT (Ro INTERCEPT) (SLOPE) ./ V
(ACTUAL Ro) 1/ V./
REICH & HAKIM
1972 .974 .780 .911 .P'
PECK & ZIERDT
1973 .956 .670 1.18
1
0.2 10 30 50 70 90 98 99.8
LAWSON
1974·1984 .950 .775 1.09 CUMULATIVE PEIICENT

PECK
1986 .986 .982 1.03 Figure 9 Distributions of deviations from the model
for the models of Figure 7, for stresses
greater than 85/85.
Figure 7 Regression par~eters for four humidity
test models.

This Appendix, then, compares the following models, REFERENCES

1. Reich and Hakim 1972
2. Peck and Zierdt 1973 1. D.S. Peck, "The Design and Evaluation of Reliable
3. Lawson 1974-1984 Plastic-Encapsulated Semiconductor Devices",
4. Peck 1986 Int'l ReI Phys Symp, 1970, pp 81-93

Figure 7 shows the regression parameters for the 2. J. L. Flood, "Reliability Aspects of Plastic En-
61 points of data collected. Figures 8 and 9 show the capsulated Integrated Circuits", Int'l ReI Phys
distributions of the deviations of the observed data Symp, 1972, pp 95-99
from each model, as numbered above.
3. M.C. Halleck, "The Ie Plastic Package: A Simple
Since all the models except 4 evolve from limited Method for Predicting Package PerforDIance" ,
data, it is not surprising that they show regions where Int'l ReI Phys Symp, 1972, pp 88-94
the model is relatively far from observed fact,
although none are off by extreme ratios. For example, 4. D.S. Peck and C.H. Zierdt, Jr., "Temperature-
the Reich and Haki. model evolved largely from the Humidity Acceleration of Metal-Electrolysis
Panama tests in high humidity, and the model fits well Failure in Semiconductor Devices", Int'l ReI
in that area, but does not provide as good an extrapo- Phys Symp, 1973, pp 146-152
49
5. N. Lycoudes, "The Reliability of Plastic Micro-
-
circuits in Moist Environments", Solid State
Technology, 1978 Oct

6. J.E. Gunn, S.K. Malik, and P.M. Mazumdar,

"Highly Accelerated Temperature and Humidity
Stress Test Technique (HAST)", Int'l ReI Phys
Symp, 1981, pp 48-51

7. R.P. Merrett, J.P. Bryant and R. Studd, "An Ap-

praisal of High Temperature Humidity Stress for
Assessing Plastic Encapsulated Semiconductor Com-
ponents", lnt'l ReI Phys Symp, 1983, pp 73-82

8. J.E. Gunn, R.E. Camenga and S.K. Malik, "Rapid

Assessment of the Humidity Dependence of IC Fail-
ure Modes by Use of HAST", Int'l ReI Phys Symp,
1983, pp 66-72

9. P. Slota, Jr. and P.J. Levitz, "A Model for a New

Failure Mechanism - Nichrome Corrosion in Plastic
Encapsulated PROM's", Proc 1983 ECC, pp 232-236

10. K. Ogawa, J. Suzuki and K. Sano, "Automatically

Controlled 2-Vessel Pressure Cooker Test Equip-
ment", IEEE Trans on ReI, v R-32, 1983 Je

11. S.P. Sim and R.W. Lawson, "The Influence of Plas-

tic Encapsulants and Passivation Layers on the
Corrosion of Thin Aluminum Films Subjected to Hu-
midity Stress", Int' 1 ReI Phys Symp, 1979,
pp 103--112

12. Private Communication, W. Gerling, Siemens AG,

Munich
13. N. Hosoya, Hirayama Mfg Corp, Tokyo, Electronic
Components and Pressure Cooker Test", publication
unknown
14. K. Ogawa, J.Suzuki and K. Sano, "Reliability Eval-
uation of Plastic Encapsulated IC's Using a New
Pressure Cooker Test", Proc, Int'l Symp for Test-
ing and Failure Analysis, 1981, pp 75-80

15. D.S. Peck, "Reliabilty of Beam-Lead Sealed-Junc-

tion Devices", Proc, Symp on Reliability, 1969,
pp 191-201

16. N. Vanderkooi and M.N. Riddell, "Dynamic Permea-

bil i ty Method for Epoxy Encapsulation Resins",
Int'l ReI Phys Symp, 1976, pp 219-222

17. H. Koelmans, "Metallization Corrosion in Silicon

Devices by Moisture-Induced Electrolysis", Int'l
ReI Phys Symp, 1974, pp 168-171

18. O. Hallberg, "Acceleration Factors for Tempera-

ture-Humidity Testing of AI-Metallized Semiconduc-
tors", SINTOM, 1979, Kopenhagen, Denmark

19. ACCELERATED TESTING HANDBOOK, D.S. Peck and 0.0.

Trapp, Technology Associates, 51 Hillbrook Dr.,
Portola Valley, CA 94025

20. R. W. Larson, "The Accelerated Testing of Plastic

Encapsulated Semiconductor Components", Int'l ReI
Phys Symp, 1974, pp 243-249

21. R.W. Larson, "A Review of the status of Plastic

Encapsulated Semiconductor Component Reliability",
Br. Telecom Technol J, vol 2 No 2, April, 1984

22. B. Reich and E. Hakim, "Envirorunental Factors Gov-

erning Field Reliability of Plastic Transistors
and Integrated Circuits". Int'l ReI Phys Symp,
1972, pp 82-87
50