INSY 7380 Accelerated Life Testing (ALT) Maghsoodloo

In most instances component reliability is so high that placing even n = 100 units on test
may not yield any failures for a test duration, of say, more than 5000 hours. If a new product is
being developed, such long testing times cannot be tolerated because the new product has to get
to the global market in due time, or else lack of market share may occur. In such cases, the
experimenter has no choice but to use accelerated testing procedures to induce failures in order
to estimate component TTF (or reliability).
Accelerated life testing (ALT), in combination with DOE (design of experiments), is
conducted by subjecting n identical units to stresses well beyond what the units on test will
experience under normal operating conditions. Such high stresses for ALT that accelerate
failure mechanism may be applied in many forms: very high, or very low temperatures,
humidity levels well beyond normal operating conditions, excessive usage, very high levels of
voltage, extreme cycling between low and high levels beyond what is considered normal
operating conditions, excessive force, excessive vibrations, ten times more units on test than
needed, etc, etc.
As Elsayed (Reliability Engineering, E.A. Elsayed, Chapter 6, Wiley INC. (2012) points
out in the beginning of section 6.4, the underlying assumption is that the failure mechanism
under ALT conditions is, except for a multiplicative factor, similar to failure mechanism under
normal operating conditions. Put differently, ALT is based on the principle that a unit under
accelerated test will exhibit the same behavioral statistical pattern in a short testing time under
very high stresses as it will exhibit in a much longer time at normal operating stresses. For
example, if the underlying failure distribution is W(, , ), then under ALT the change in the
shape parameter  will be much smaller than the changes in minimum life  and scale parameter
  . That is to say, under ALT the change in  (or overall process variability, or CV) will be
negligible compared to changes in  and the characteristic life  as compared to normal
operating conditions. There are 3 different physical models that have been developed in the past
115 years that can be used to estimate the MTTF under normal operating conditions (o = normal
operating conditions) from ALT data, where subscript “s” will be used to designate statistics
computed under high stressed conditions.

213
(1) The Arrhenius Model
This is the most commonly used model relating TTF to high thermal stresses. Thermal
stresses occur in solid state diffusion, chemical reactions, many semiconductor failure
mechanism, battery life, etc. The underlying distribution of TTFo (TTF under normal operating
conditions) in almost all these cases is exponential, Weibull, or lognormal (i.e., all positively
skewed pdfs). The Arrhenius rate law that describes the (failure) rate, r, at which reaction to
temperature of the test unit occurs is given below.
 Ea /(k Τ   B/ Τ
r = C e = C e (102)

where C is a constant which is characteristic of the failure mechanism of the item under test, Ea
= the activation energy needed to induce failure measured in eV (= electron volts; close to
vaporization energy for metals and chemical bond energies for polymers), k = the Boltzman’s
constant = 8.6171105 eV/Kelvin (Note that Google.com gives k = 8.616105, while L. W.
Condra, p. 232, gives k = 8.617105 ) and T = the temperature in Kelvin = Centigrade +
273.15, and B = Ea/k. In RE engineering, the Arrhenius model is also used to measure the
impact of temperature on reliability because we make the assumption that the TTF is inversely
proportional to the reaction (failure) rate, r, given in equation (102), i.e.,
Ea /(k T ) B/ Τ
TTF = C e = Ce (103)
where C  1/C is the constant of proportionality characteristic of the product under test. The
Arrhenius model is applicable when the product r1TTF1 = r2 TTF2, where r1 and r2 are

reaction rates at testing temperatures T1 and T2, respectively. The relationship Rate1TTF1 =
Rate2 TTF2 implies that r TTF will practically stay constant over the range of temperature

applicability, and as a result roTTFo  rsTTFs, where TTFo represents TTFo under normal

operating conditions and TTFs represent component life under (accelerated) stressed conditions.
Thus,
1 1 1 1
TTFo rs C e  Ea /(k Ts ) Ea (
To

Ts
)/k Ea ( 
To Ts
)/k
= = 
=e  TTFo = TTFs e
TTFs ro C e aE /(k To )

(104a)
Eq. (125a) shows that the acceleration factor for the Arrhenius model is given by

214
 1 1 1 1
Ea (  )/k B(  )
To Ts To Ts
Af = TTFo / TTFs = e =e (104b)
Note that the smaller the required activation energy Ea is, the more rapidly the unit on test will
fail resulting in smaller Af value.

The Example 6.11 on page 405 of E. A. Elsayed. In this example n microelectronic devices
(the value of n not specified by the author) are put on accelerated test at Ts = 200 Celsius = 200

+ 273.15 = 473.15 Kelvin and the MTTfs of the n units was approximately equal to 4000 hours.
The operating temperature To = 50 C = 323.15K, and the required activation energy was 0.191
1 1 1 1
Ea (  )/k 0.191(  )/8.6171105
To Ts 323.15 473.15
eV. Thus, the sample MTTfo = MTTfs e = 4000 e =

35191.33024 hours. The value of acceleration factor is Af = 35191.33024 /4000 = 8.79783256.

Example 16. The TTF of n = 10 samples under an accelerated temperature of Ts = 100

Centigrade are t(i) : 130, 140, 160, 180, 185, 195, 205, 205, 240, and 260 hours. The
measurement of interest is the thermo-compression bond between two dissimilar metals, the
strength of which reduces in time by the formation of voids by solid-state diffusion which has
an activation energy of 0.90 eV. The normal operating temperature is To = 25 Celsius. The

sample statistics are x s =190, Ss = 40.8248290, and cvs = 21.487% showing that the accelerated

data is obviously not exponentially (i.e., IFR) and if it is Weibull, then the slope   5.0 (see my
Table 1 on p. 10). Most probably, the accelerated data is lognormally distributed. The use of
equation (104a) yields the normal operating sample mean to failure mttfo = 190 
1 1
0.90(  )/(8.6171105 )
e 298.15 373.15 = 1901142.3450161 = 217045.5531 hours  Af =
1142.3450161. If we wish to be more conservative about our estimate of MTTF in normal
1 1
0.90(  )105 /8.6171
operating use, we could estimate it as mttfo = 130 e 298.15 373.15 = 148504.8521

hours, giving an acceleration factor of Af = 148504.8521/190 =781.6044847, where 130 is the

value of the 1st order statistic, x(1) = t1, under stressed condition. Note that we are using mttf as
the sample MTT failure.
It is reported in the literature that the value of Ea ranges in the interval 0.30  0.60 for

215
semiconductor failures, for intermetallic diffusion (like in Example 16) it ranges in the interval
0.90 1.1 eV, and for silicon junction defects Ea = 0.80 eV. The question arises how high the
stressed temperature should be for a unit under accelerated test so that the resulting stressed life
can be extrapolated to the expected life under normal operating conditions. Almost all metals
change properties when the testing temperature exceeds 50% of their melting temperature Tm.
Therefore, the accelerated testing temperature, Ts, must not exceed 0.50Tm.

Example 17. The lifetimes of n = 50 PC components under an accelerated temperature

of Ts = 100C yielded the sample mean x s = 232.2 hours and a standard deviation of Ss = 82.8

hours, with Ea = 0.85 eV and To = 27C. The use of equation (104b) gives an acceleration
1 1
Ea (  )/k
factor of Af = e 300.15 373.15 = 619.695651 giving an estimated mttfo = Af  x s =

619.695651232.2 = 143893.3301 hours  16.42618 years. Since the sample size n > 20, then
we may use the normal approximation to the SMD of x s to obtain an approximate lower 95%

CI for the MTTFs, given by Ls = 232.2  1.64582.8 / n = 212.937563 hours  Lo =

619.695651212.938 = 131956.75253264 hours  15.0635562252 years  MTTFo <  at the

95% confidence level. Note that this normal approximation would not be permissible unless n >
20. Methods of analysis for the exponential, Weibull, and lognormal underlying distribution of
TTFs, for any n, are given by Wayne Nelson, (1990), “ Accelerated Testing, Wiley, New York,
ISBN: 0-471-52277-5.

Determination of the Acceleration Factor Af Using Linear Regression

In order to use regression to estimate Af, the Arrhenius model must first be linearized as
Ea / (k T)
shown below. From equation (103), TTF = C e , which can be linearized by taking the
natural logarithm of both sides only once. This leads to y = ln(TTF) = ln (C) + Ea /(kT) =
ln (C) + Ea x , where x = 105 /(8.6171T), and T must be in units of Kelvin. I used the data
provided by Boris Gnedenko et al (Statistical Reliability Engineering, Wiley, Example 5.2 on
pp. 171-172, ISBN: 0-471-12356-0) and W. Nelson (1990), which are listed for your
convenience below, to estimate C and Ea using regression analysis. The experiment from the

216
 above two authors involved a new Class-H motor insulation with a design temperature of To
= 180 C = 453.15 Kelvin, where n = 40 units were equally and randomly divided and tested
to failure at the accelerated temperatures 190, 220, 240, and 260 Celsius. The accelerated
times TFs in hours are provided in Table 5.2 of B. Gnedenko and duplicated herein. I used
Minitab to regress y on x, where x = 105/(8.6171T), and its output is provided below.

Table 5.2 of Boris Gnedenko et al (page 171)

190 C 220 C 240 C 260 C
7228 hours 1764 hours 1175 hours 600 hours
7228 2436 1175 744
7228 2436 1521 744
8448 2436 1569 744
9167 2436 1617 912
9167 2436 1665 1128
9167 3108 1665 1320
9167 3108 1713 1464
10511 3108 1761 1608
10511 3108 1953 1896
mttf = 8782.20 2637.6000 1581.4000 1116.0000
S = 1244.0117 453.5654 244.2745 439.2357
cv = 14.165% cv = 17.196% cv = 15.447% cv = 39.358%
Regression Analysis: y versus x

The regression equation is

y = - 7.28341 + 0.64936 x, x = 105/(8.6171 T)

Predictor Coef SE Coef T P

Constant -7.2837 0.7719 -9.44 0.000
x 0.64938 0.03317 19.58 0.000

S = 0.2557 R-Sq = 91.0% R-Sq(adj) = 90.7%

Analysis of Variance
Source DF SS MS F P
Regression 1 25.073 25.073 383.36 0.000
Residual Error 38 2.485 0.065
Lack of Fit 2 0.368 0.184 3.12 0.056
Pure Error 36 2.118 0.059
Total 39 27.558

217
Unusual Observations
Obs x y Fit SE Fit Residual St Resid
11 23.5 7.4753 7.9973 0.0416 -0.5220 -2.07R
39 21.8 7.3827 6.8509 0.0635 0.5319 2.15R
40 21.8 7.5475 6.8509 0.0635 0.6966 2.81R

R denotes an observation with a large standardized residual

The above regression output clearly shows that ŷ =  7.28341 + 0.64936 x is an excellent
2
model because R Model = 91% so that ln(C) =  7.28341  C = 0.00068684 and Ea =

0.649358eV. To extrapolate the expected life to the operating temperature of 180C = 453.15
K, we insert xO = 100000/(8.6171453.15) = 25.609251 into our regression model  ŷO =

 7.28341 + 0.649358xO = 9.34617  mttf o = e9.34617 = 11459.8130 hours. I will next

convert the above regression model ŷ =  7.28341 + 0.64936x , ln(TTF) = ln (C) + Ea /(k T) to
the Arrhenius format:
5 / (8.6171T)
mttf o = 0.00068684 e0.64936×10 , (105)

where 0.00068684 = eC = e  7.28341 and the temperature T must be measured in Kelvin.

Inserting To = 180 + 273.15 = 453.15 into equation (105) again yields mttfO (180C) =

11459.8130 hours. The acceleration factors from 180 to 190C is Af = 11459.8130 /8782.20 =
1.3043. In practice, I would use only the stressed-Temperature that is closest to TO to compute
the AF. Further, I attempted to improve the above model by adding the regressors x2 and x3 to
2
the model, unfortunately the R Model improved by a minute amount to 92.3% but all the
coefficients in the model became highly insignificant (i.e., a worthless model).

Exercise 26. Boris Gnedenko et al mention on their page 172 that the failure data at
260 Celsius in the above experiment looks very suspicious because it exhibits much higher
sample cv than the other 3 accelerated temperatures. That is to say, the failure mechanism at
260 C is different from failure modes at lower temperatures. Repeat my analysis of the above
experiment but remove the data at 260 C. ANS: mttf o (180 C) > 12000 hours.

218
(2) The Inverse Power Law (IPL)
This law is generally used when the TTF is inversely proportional to the applied
(accelerated) stress, and the underlying lifetime distribution is almost always Weibull, or
perhaps lognormal. As in the case of Arrhenius model, the IPL model is applicable only when
there is a single type of stress, which in most cases is voltage accelerated stress, alternating
temperature stress, or mechanical vibration in order to induce fatigue failure. The general form
of the IPL is given by
TTFs = C/ Sb (106a)

where C > 0 and the exponent b > 0 are constants characteristics of the items under test, and S
is the applied (accelerated) stress. The value of the exponent b = [2, 3] for metals and
electronic solder joints, b = [4, 10] for microelectronic parts, and b = [4, 7] for intermetallic
fatigue failures.

Example 22 (borrowed from L. W. Condra, RE Improvement with DOE, pp 236-237,

Marcel Dekker, ISBN: 0-8247-0527-0). A sample of n electronic solder joints are placed on
accelerated fatigue-testing at a displacement of S = 0.0008 inches with a MTTfs = 10 cycles.
Assuming that under normal use the maximum displacement is So = 0.00005 inches and the
exponent b = 2.50, our objective is to estimate MTTFo. We need to compute the value of the Af

C/ Sbo
= Sb/ So = (0.0008/0.00005)2.5 = 1024  mttfo = 102410 =
b
= mttfo/ MTTfs =
C/ S b

10,240 cycles to failure.

Elsayed provides another form of IPL given in his equation (6.67) which is modified as
TTFs = C(Vo / Vs)b (106b)

where Vo is the standard specified (voltage) operating stress, Vs is the accelerated voltage

stress, and the constant C is characteristic of the product under test , fabrication method, etc.

The Example 6.13 on pages 410-411 of Elsayed. In this experiment two samples of 20

CMOS integrated circuits each are put on accelerated life test, where Vs represents accelerated

electric field stresses at 10 and 25 eV. The underlying distribution of TTF is assumed Weibull
and there is only one stress factor, namely electric field, and hence the IPL is a plausible model

219
for TTFs. The normal operating stress is at Vo = 5 eV. For your convenience I have duplicated
Table 6.7 of E. A. Elsayed on his page 410 below. I first used the data under the two
accelerated stress levels, V1 = 10 and V2 = 25 eV, to obtain the MLEs of the Weibull

parameters  and . Using my equations 111 (a &b) the MLEs are ˆ 1s = 1.836028, ˆ 1s =

9343.5856011 hours, and at V2 = 25 eV, ˆ 2s = 1.981834234, ̂ 2s = 3916.9661061541

hours. These MLEs under stressed conditions are consistent with those of Elsayed’s. It seems
that if the stressed data set is W(0, , ), then a rough value of the Weibull slope is close to  
1.910. However, it is not clear what the estimate of the characteristic life at normal operating

stress 5 eV is, because ˆ 1s = 9343.585601 hours and ˆ 2s = 3916.9661061541 hours were

obtained under accelerated testing conditions. I will now obtain the least-squares estimate of .
Table 6.7 of Elsayed page 410 (TTF under accelerated testing condition)
10 eV 1037.39 hours, 3218.11, 3407.17, 3520.36, 3879.49, 3946.45, 6635.54, 6941.07, 7849.78, 8452.49

10 eV 9003.08, 9124.50, 9365.93, 9642.53, 10429.50, 10470.60, 11162.90, 12204.50, 12476.90, 23198.30

25 eV 809.10, 1135.93, 1151.03, 1156.17, 1796.53, 1961.23, 2366.54, 2916.91, 3013.68, 3038.61

25 eV 3802.88, 3944.15, 4095.62, 4144.03, 4305.32, 4630.58, 4720.63, 6265.99, 6916.16, 7113.82 hours

In order to obtain a LS estimate of , I first linearized the IPL model, TTFs = C/ Sb, by taking

the natural logarithm of both sides. This leads to y = ln(TTFs) = ln(C)  b ln(S) , where S is at

2 levels, 10 and 25 eV. I used Minitab to regress y = ln(TTFs) on x = ln(S), with the following
output.

The regression equation is

y = 11.0 - 0.941 x, x = ln(S)

Predictor Coef SE Coef T P

Constant 11.0065 0.6383 17.24 0.000
x -0.9406 0.2281 -4.12 0.000

S = 0.6609 R-Sq = 30.9% R-Sq(adj) = 29.1%

Analysis of Variance

Source DF SS MS F P
Regression 1 7.4281 7.4281 17.01 0.000
Residual Error 38 16.5968 0.4368
Total 39 24.0248

220
Unusual Observations
Obs x y Fit SE Fit Residual St Resid
1 2.30 6.979 8.841 0.148 -1.862 -2.89R

I must caution the reader before using the above Minitab model for extrapolation! You must
2
observe that the value of R Model = 30.9% is woefully too small to be an adequate model due
to the fact that there is too much within (or experimental error) variability in the data. The data
under level 1 of stress (10 eV) ranges from 1037.39 stressed hours to 23198.30 hours, which is
very large, but still the regression is highly significant. The above model cannot be improved
because there are only 2 levels of stress factor and hence regression can have only one df and
any attempt to improve it by adding regressors such as x2, x3 and 1/x to the model will be futile
because the design does not provide but one df for studying effects. Hence, we have to
2
extrapolate with a model whose R Model = 30.9%. The estimate of the constant Ĉ = e11.0065 =

60264.591, and the estimate of the exponent b̂ = 0.9406 is very close to Elsayed’s answer of
0.95318. To estimate the MTTFo at 5 eV, we insert x = ln(5) =1.609438 into the regression
model. This yields ŷ (1.60944) = 11  0.94061.60944 = 9.49266  mttf o = e9.49266 =

13262.06124 hours, which is fairly close to Elsayed’s answer of 12140. Since the Weibull
1 1
mean E(T) =  Γ(1 + ) , then θ̂o = 13262.06124 / Γ(1 + ) = 14947.92193 hours. The
β 1.910

two acceleration factors are Af1 = 13262.06124 /8298.3295 = 1.59816036, where 8298.3295 =

mttf(at 10 eV) and Af2 = 13262.06124 /3464.2455 = 3.82827.

(3) The Eyring Model

Both the Arrhenius and IPL models include only the effect of one accelerated
stress. The Eyring model contains two stress factors, one of which is always temperature
stress, and the other can be any stress type such as electric field, voltage, humidity, mechanical
stress (load per area), even temperature cycling, or electrical current stress. The rate of reaction
(or rate of failure) to the two stresses is given by
 Ea /(k 
r = C1 e Sb (107a)

221
where r is the rate of reaction to the two stresses; r may be thought of the parameter  if the

underlying distribution is exponential, but r  1/  if the TTF is W(0, , ), and if TTF is

lognormal, then r = 1/T0.50 (the inverse of median life). Thus, from (107a) we deduce that
Ea /(k Τs 
TTFs = C e / Sb = C eEa /(k Τs  Sb (107b)

The values of Ea and exponent b can be obtained empirically once accelerated data are available.
For electronic applications, b  2 to 3 and Ea = 0.90 eV, and C is a constant characteristic of the
product and testing conditions. Equation (107b) implies that under normal operating conditions
the TTF is given by
Ea /(k o 
TTFo = C e / Sbo (107c)
Combining equations (128 b&c) yields

TTFo eEa /(k Τo  / Sbo (E / k)(1 / To - 1 / Ts)

Af = = E /(k Τ  = (S/So)b e a (108)
TTFs e a s /S b

Note that Af must be directly proportional to Ea because larger activation energy required to
induce failure in the test unit generally implies longer MTTFo.

Example 18. L. W. Condra (RE Improvement with DOE, 2nd edition, Marcel
Dekker) reports (on his p. 239) the results of an accelerated life testing experiment of n
(unspecified) microelectronic circuits conducted at the standard accelerated temperature stress
of 85 C and a standard accelerated relative humidity (RH) of S = 85%. (He refers to this type

of accelerated testing as Temperature-Humidity Operating Bias test.) The sample MTTfs is

reported to be 800 hours and normal operating conditions are To = 40 C and RHo = 60%. The

Model (107b) when the 2nd stress represents S = RH (relative humidity) is referred to as Peck’s
relationship. Peck, D. S. (1986) ”Proc. International RE Physics Symposium, 24, pp. 44-45,
reports an exponent value of b  2.70 and an activation energy of Ea = 0.79 eV, but Condra in
his example uses the rough values of b = 3 and Ea = 0.90 eV. I will use Peck’s values in
equation (129) to estimate the acceleration factor Af.

222
 (0.79105 /8.6171)(1/313.16 - 1/358.16)
Af = (85/60)2.7 e = 101.3547
which yields mttfO = 101.3547 800 hours = 81083.74783 hours = 9.2561356 years. The above
estimated value of mttfO = 9.2561356 years does not conform well with that of Condra’s 16.6
years. If we use Condra’s values of b = 3 and Ea = 0.90 eV in equation (108), we obtain Af =
187.780224 and an estimated mttfO = 187.780224  800 = 150224.179364 hours = 17.148879
years. I tried to obtain Condra’s answer of Af = 182 by using his values of To = 313 and Ts
= 358 in equation (108) but I still got an answer of Af = 188.5450005 which is not equal to
Condra’s answer of 182. The reader should be careful about interpreting the values of mttfO
because if the underlying distribution is exponential, then mttfO is an estimate of MTTF; if the
underlying distribution is Weibull, then mttfO is an estimate of the characteristic life tc = , and
if the underlying distribution is lognormal, then mttfO is an estimate of the median life.
Furthermore, the farther the operating conditions are from the stressed conditions, the less
accurate the regression estimates of b and Ea become. This problem gets compounded when the
baseline distribution is very highly skewed and /or there are outliers in the data.

Example 6.14 on page 412 of Elsayed. The data listed in Table 6.8 on page 412 of Elsayed
presents the results of an ALT with 8 FLCs (factor level combinations) of Temperature and
Voltage stresses. For your convenience, I am providing Elsayed’s data below. The normal
operating temperature To = 30 C = 303.16 Kelvin and the operating voltage is Vo = 25 volts.
Instead of using Elsayed’s parametric approach to estimate MTTFo, I linearized the Eyring
Model (107b) as follows: y = ln(TTFs) = ln ( C)  b ln(Vs) + Ea x, where x = 100000/(8.6171
Ts), and then I regressed y on ln(Vs) and x. The Minitab output is given below
The regression equation is
y = 2.25333 - 0.426737 LV + 0.199714 x, LV = ln(Vs), x = 100000/(8.6171Ts),

Predictor Coef SE Coef T P

Constant 2.253 1.405 1.60 0.170
LV -0.42674 0.03955 -10.79 0.000
x 0.19973 0.04057 4.92 0.004

S = 0.05824 R-Sq = 96.6% R-Sq(adj) = 95.2%

Analysis of Variance

Source DF SS MS F P
Regression 2 0.47708 0.23854 70.32 0.000

223
Residual Error 5 0.01696 0.00339
Total 7 0.49404

Source DF Seq SS
LV 1 0.39488
x 1 0.08220

Table 6.8 of E. A. Elsayed (p. 412)

Voltage 50 100 150 200 volts

Temperature
60 C 1800 hours 1500 1200 1000
70 C 1500 1200 1000 800 hours

In the above Minitab output, LV = ln(Vs) and y = ln(TTFs). Note that the value of Ea from the
above regression output is Ea = 0.199714 which is in agreement with the reported value by
Elsayed on his page 412. In order to estimate the MTTFo, I used extrapolation (which is
generally not a good idea in regression analysis) in the above regression model, which has an
2
excellent R Model , as follows: ŷ(30 oC, 25 volts) = 2.25333  0.426737ln(25) + 0.199714 

100000/(8.6171303.15) = 8.52494  mttfO = e8.52494 = 5038.8636 hours, which is a bit larger

than Lo = 4484.11 hours reported by Elsayed. There will be 8 different values of Af because
there are 8 FLCs of the two stresses. The value of Af from normal operating conditions (30C,

25 volts) to stress FLC (60 C, 50 volts) is Âf = 5038.8636 /1800 = 2.79937. To verify the

adequacy of the Eyring model to the data, we also need to estimate this last acceleration factor
50 0.426737 (19971.4/8.6171)(1/303.15  1/333.15)
Af from equation (108) as follows: Â f (Model) = ( ) e
25
= 2.6758, which is fairly consistent with the regression-value of 2.7994.

Example 6.7 of Elsayed on pages 388-389. This experiment makes no assumptions about
the underlying distributions of Times TF (i.e., the nonparametric) and uses regression to
estimate the MTTF by extrapolation. I used the data in Table 6.1 of Elsayed on his page 388 to
regress the TTF on stress factor Temperature in Kelvin, and stress factor electric field measured
in units of eV. For your convenience, I am duplicating Elsayed’s Table 6.1 on the next page.

224
The resulting Minitab output is given below:
Regression Equation

TTFs = 6061.97 - 17.8487 T + 160.159 eV

Coefficients
Term Coef SE Coef T P
Constant 6061.97 3.55841 1703.56 0.000
T -17.85 0.01343 -1329.01 0.000
eV 160.16 0.22473 712.68 0.000

Table 6.1 (on page 388 of Elsayed)

Temperature C 100 100 100 100 100 100
Electric Field (eV) 10 10 10 10 10 10
Stressed TTF 1000 hours 1002 1003 1004 1005 1006 hours
(TTFs)

Temperature C 150 150 150 150 150 150

Electric Field (eV) 10 10 10 10 10 10
Stressed TTF 110 110.5 110.7 111 111.4 111.8 hours
(TTFs)

Temperature C 200 200 200 200 200 200 200 200 200 200
Electric Field (eV) 15 15 15 15 15 15 15 15 15 15
Stressed TTF 19 19 19.1 19.2 19.3 19.32 19.38 19.4 19.44 19.49
(TTFs)

Summary of Model

S = 1.16308 R-Sq = 100.00% R-Sq(adj) = 100.00%

PRESS = 36.9518 R-Sq(pred) = 100.00%

Analysis of Variance

Source DF Seq SS Adj SS Adj MS F P

Regression 2 3967238 3967238 1983619 1466367 0
T 1 3280158 2389312 2389312 1766270 0
eV 1 687080 687080 687080 507916 0
Error 19 26 26 1
Total 21 3967264

Fits and Diagnostics for Unusual Observations

Obs TTFs Fit SE Fit Residual St Resid

17 1000 1003.33 0.474824 -3.33333 -3.13951 R

225
 22 1006 1003.33 0.474824 2.66667 2.51161 R

R denotes an observation with a large standardized residual.

Predicted Values for New Observations

New Obs Fit SE Fit 95% CI 95% PI

1 1541.19 0.836013 (1539.44, 1542.94) (1538.19, 1544.18)

Values of Predictors for New Observations

New Obs T eV
1 298.15 5 X

X denotes a point that is an outlier in the predictors.

General Regression Analysis: TTFs versus TempC, eV ; using the centigrade data

Regression Equation

TTFs = 1186.61 - 17.8487 TempC + 160.159 eV

Coefficients

Term Coef SE Coef T P

Constant 1186.61 1.29166 918.67 0.000
TempC -17.85 0.01343 -1329.01 0.000
eV 160.16 0.22473 712.68 0.000

Summary of Model

S = 1.16308 R-Sq = 100.00% R-Sq(adj) = 100.00%

PRESS = 36.9518 R-Sq(pred) = 100.00%

Analysis of Variance

Source DF Seq SS Adj SS Adj MS F P

Regression 2 3967238 3967238 1983619 1466367 0
TempC 1 3280158 2389312 2389312 1766270 0
eV 1 687080 687080 687080 507916 0
Error 19 26 26 1
Total 21 3967264

Fits and Diagnostics for Unusual Observations

Obs TTFs Fit SE Fit Residual St Resid

17 1000 1003.33 0.474824 -3.33333 -3.13951 R
22 1006 1003.33 0.474824 2.66667 2.51161 R

R denotes an observation with a large standardized residual.

Predicted Values for New Observations

New Obs Fit SE Fit 95% CI 95% PI

1 1541.19 0.836013 (1539.44, 1542.94) (1538.19, 1544.18)

Values of Predictors for New Observations

New Obs TempC eV

226
 1 25 5 X

X denotes a point that is an outlier in the predictors.

General Regression Analysis: LnTTF versus x, LV

Regression Equation

LnTTF = -11.6518 + 0.599347 x - 0.0332739 LV

Coefficients

Term Coef SE Coef T P

Constant -11.6518 0.0659043 -176.798 0.000
x 0.5993 0.0011245 532.968 0.000
LV -0.0333 0.0151534 -2.196 0.041

Summary of Model

S = 0.00715762 R-Sq = 100.00% R-Sq(adj) = 100.00%

PRESS = 0.00124123 R-Sq(pred) = 100.00%

Analysis of Variance

Source DF Seq SS Adj SS Adj MS F P

Regression 2 58.9100 58.9100 29.4550 574939 0.0000000
x 1 58.9098 14.5526 14.5526 284055 0.0000000
LV 1 0.0002 0.0002 0.0002 5 0.0407246
Error 19 0.0010 0.0010 0.0001
Total 21 58.9110

Fits and Diagnostics for Unusual Observations

Obs LnTTF Fit SE Fit Residual St Resid

1 2.94444 2.95815 0.0022634 -0.0137082 -2.01878 R
2 2.94444 2.95815 0.0022634 -0.0137082 -2.01878 R

R denotes an observation with a large standardized residual.

In order to estimate the MTTFo from the above regression models at the normal operating
temperature To = 25C = 298.15 Kelvin, and 5 eV, we insert these values into the first model as
follows: mttfO = 17.8487298.15 + 160.1595 = 1541.19 hours which is consistent with that
of Elsayed’s answer. The value of the acceleration factor from normal operating conditions

(25C, 5 eV) to stress levels (100C, 10 eV) is equal to Âf1 = 1541.19/ mttf(100 o C,10 eV) =

1541.19/1003.33333 = 1.5331. Equation (108) gives Âf1 (Model) =

10 0.0332739 (59937.7/8.6171)(1/298.15  1/373.15)

( ) e = 2.367, which is not consistent with  f1 =
5
1.533, which may warrant the rejection of the Eyring model.

227
Chapter Summary
1. The acceleration factor For the Arrhenius Model is given by Af =

1 1
Ea (  )/k
To Ts
e  MTTFo = Af MTTFS, where k = Boltzman’s constant =
8.617110 5. Two cases exit: (a) The required activation energy Ea to induce
failure is known, (b) Ea is not known and has to be empirically estimated from
accelerated data. For Semiconductor failure 0.30  Ea  0.60; for intermetallic
diffusion 0.90  Ea  1.10; For silicon junction defects Ea = 0.80.

(a) Assume Ea = 0.50 and normal operating temperature is 25 C and accelerated

testing is done at 50 C. Then To = 25 + 273.15 = 298.16K and TS = 50 + 273.15 =
323.1600K 
1 1
0.50(  )105 / 8.6171
Af = e 298.15 323.15 = 4.506862
Note that Af is an increasing function of Ea because larger values of Ea imply that
more energy is required to induce failure which in turn would lead to higher
MTTFO. Note that some sources use the conversion Kelvin = C + 273.15 and
others use Kelvin = C + 273.16.

(b) Ea is unknown.
Identify at least two stressed temperature levels, such as 50C and 75C (<
0.50Tm ) and obtain stressed failure data. Linearize the Arrhenius model TTFS =
E /(k Ts )
Ce a and regress ln(TTFS) on x = 105 /(8.6171T); then the rough estimate
of Ea is given by the slope of the regression line. However, one must be
cognizant of the fact that extrapolation is classical regression violates
regression assumptions and is generally frowned upon. But than when there are
no information about Ea (physical or otherwise), then the regression approach
would be the only way to obtain a statistically unsound manner of obtaining a
rough estimate of the activation energy Ea.

228
2. The IPL : TTFs = C/ Sb  Larger values of b induce
higher failure rate reaction and smaller TTF. The value of
b = [2, 3] for metals and electronic solder joints, b = [4, 7] for
intermetallic fatigue failure, and b = [4, 10] for microelectronic parts, and very
rarely b lies outside the range [2, 20].
b
(a) b is known  Af = Sb/ So . For example, suppose the normal operating

voltage is So = 110V, stressed voltage is S = 220 and b = 2.8. Then Af =

(220/110)2.8 = 6.9644.

(b) b is unknown. First linearize TTFs = C/ Sb 

y = ln(TTFs), x = ln(S), y-intercept= ln(C), and b̂ =  slope of the LS line.

Ea /(k Τs  Ea /(k Τs 
3. The Eyring Model : TTFs = C e / Sb = C e Sb

Af = (S/So)b e(Ea / k)(1 / To - 1 / Ts)

(105 Ea / 8.6171)(1 / To - 1 / Ts)
=(S/So)  e
b

(a) Both Ea and b are known.

(b) At least one is unknown. Use stressed data to extrapolate to estimate b

and Ea. Note that this extrapolation often does not provide adequate and /or
reasonable estimates of Ea and b, which implies that the Eyring model does not
fit the data, and/or regression assumptions are grossly violated. Further,
extrapolation is always on poor statistical ground and is used in accelerated
testing because there are no other options, i.e., the constants b and Ea are
unknown and testing under normal operating conditions involves well over
thousands of hours.

229