C4711_C02.

fm Page 25 Saturday, February 28, 2004 10:29 AM

2
Demarcation Mapping:
Initial Design of Accelerated Tests

Accelerated testing is a common tool in reliability studies. In its classical

implementation, it works especially well with devices of intermediate reli-
ability such as InGaAs laser diodes. With accelerated testing we can uncover
the (known) failure modes and obtain precise estimates of reliability, so we
know how much redundancy to build in to make a reliable system. However,
for ultrahigh-reliability devices, the use of accelerated testing becomes more
problematic. We attempt to deduce the reliability of a device (i.e., lack of
failure or degradation) over a very long period of time at operating stress,
from a short exposure to high stress. Any degradation or failure that is
observed during the accelerated test that is relevant during operating life
disqualifies the device. Failure or degradation that represents an artificial
failure mode provides no useful information about the reliability of the
device under the conditions in question and, in fact, could mask a relevant
failure mode.
If failure and degradation do not provide information on reliability (except
in cases where an enormous amount is already known about the failure
physics of the device under test), then where is the information on reliability
in accelerated tests? The information on reliability in an accelerated test of a very
high reliability device is in the stress-time regimes free from and prior to any
degradation or failure. This concept is rather radical for most practitioners and
theorists involved in accelerated tests, but it is inescapable if one allows for
the possibility of one degradation or failure mode masking another.
For most new devices, there is some prior experience with similar devices,
hence some expected failure or degradation modes. There is also, however,
always the possibility of something new. So truly rigorous testing may
require both characterizing the acceleration of known failure modes (or
screening them out if possible), and reducing the risk from unknown failure
modes. Characterizing the acceleration of known failure modes is the stan-
dard problem in accelerated testing, so the theory in texts such as Nelson
(1990) or Meeker and Escobar (1998), possibly augmented by models and
methods such as described in Chapters 3, 5, and 6 of this book, is sufficient.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 26 Saturday, February 28, 2004 10:29 AM

However, both for the development of screens that do not reduce the remaining
life significantly, and for reducing risk of a priori unknown failure modes, it is
necessary to have in place a theory from which it is possible to infer reliability
from the initial failure and degradation-free periods in accelerated testing.
This chapter provides an introduction to some theory and methods nec-
essary for inferring reliability from failure and degradation-free periods. The
statistical theory underlying this development was begun in Complements
1.6.3 and 1.6.4. The physical theory and the integration of the two occupy
most of this chapter and Chapter 4. In this chapter we introduce the notion
of demarcation maps, both analytic (LuValle et al., 1998; LuValle, 2000) and
computational (LuValle et al., 2002). In conjunction with this development,
we show how demarcation maps can be used to accomplish the following:

1. Design and interpret acceptance tests: Acceptance tests are acceler-

ated tests where no failures during the test are meant to imply
something positive about reliability. An example of an acceptance
test can be given for the case of a device with a solder joint that has
a low (e.g., 100°C) melting temperature. The problem is to specify
the combination of time and temperature that will be at least as hard
on the device as the design lifetime exposure under service condi-
tions, with the constraint that the acceptance test temperature must
be below 100°C if it is to reflect design lifetime conditions.
2. Design and interpret safety tests for screens: Safety tests are tests
that examine the residual life of devices that survive a screen meant
to eliminate early failures.
3. Design and interpret certain step stress experiments: Here the step
stress experiments are designed to probe for failure and degradation
modes that are masked by failure or degradation modes observed
at high stresses.

Demarcation map theory allows study of one specified mechanism with

unknown parameters. Evanescent process maps, covered in Chapter 4,
allows this to be generalized to the study of neighborhoods of mechanisms.
In this chapter and in Chapter 4, both the power and ultimate limitations of
accelerated testing are exposed. One result of full understanding of both of
these chapters is that there is always some risk involved from the reliability
point of view in introducing a new material system to a new operating
environment. Understanding these two chapters gives the reliability practi-
tioner the ability to articulate that risk explicitly in terms of probability and
physical mechanism, and to systematically reduce that risk.
Demarcation maps can be divided into two categories: analytical and com-
putational. These are described below:

1. Analytical demarcation maps: These are developed by constructing an

analytical approximation to determine which among the set of poten-
tial chemical reactions with given parametric form are complete. Given

© 2004 by CRC Press, LLC

C4711_C02.fm Page 27 Saturday, February 28, 2004 10:29 AM

the approximation and the stress conditions, we can divide the set
into two regions. One region is where all the potential reactions are
essentially completed (the region of potential failure modes) and the
other where they are not. By comparing these maps for two different
stress trajectories, an accelerated stress trajectory, and an operating
trajectory, four regions are identified: (a) a region where the potential
reactions are not complete for either stress condition, (b) another
where they are complete for accelerated test conditions but not for
operating life conditions (the region of potential artificially generated
failure modes), (c) another where they are complete for operating life
but not for accelerated test conditions (a region of potential failure
modes not uncovered by the accelerated test), and (d) a region where
they are complete for both stress conditions.
2. Computational demarcation maps: The concept is the same as that of
the analytical, but rather than make analytical approximations to
determine when reactions are complete, the extent of potential reac-
tions is calculated for each stress trajectory and compared directly
at each point. The dominating trajectory is indicated at that point
by symbol or color choices.

The latter is, in fact, more flexible and lends itself to both more complex
kinetic processes, and more complex statistical analysis. However, it also
requires more sophisticated software for doing the calculations. The former
is more easily calculated, and offers surprisingly useful insights. We will begin
this chapter studying the tools and theory of analytical demarcation maps.
Although at present it seems impossible to make a detailed map of this
sort for all chemical reactions, it is possible to construct a fairly conservative
map for a moderate-sized set of chemical reactions (LuValle et al., 1998, 2000).
The theory supporting the conservative nature is fairly involved, but the
basic idea is straightforward and is provided in the next section. Following
that are examples for several kinds of degradation processes, listed below.
The first three are based on the assumption of simple, one-step processes or
empirical models. The fourth example uses computational demarcation map-
ping with more complicated kinetic processes.
For simplicity, in the first seven sections of this chapter, where the physical
theory of demarcation mapping is developed, we ignore the question of
sample size and statistical interpretation. We return to statistical theory in
Sections 2.8 and 2.9.

2.1 Analytical Theory of Thermal Demarcation Maps

The demarcation approximation was originally developed in the context of
material systems with a macroscopically observable response driven by a

© 2004 by CRC Press, LLC

C4711_C02.fm Page 28 Saturday, February 28, 2004 10:29 AM

set of parallel thermally activated (Arrhenius) first-order kinetic processes

(see Complement 1.6.1). It is assumed that each process consists of a single
step and there is a distribution of activation energies across processes (Pri-
mak, 1955). With these conditions the rate constant for each local process
has the form:

 −E 
k 1 = ν exp a  (2.1)
 kT 

where ν is a premultiplier, k is Boltzmann’s constant, T is absolute temper-

ature, and Ea is the activation energy (in the following, all activation energies
are given in electron volts, eV). The relative amount (normalized to total
amount that can react, also called the reaction extent) reacted over time for
this local process then is (see Complement 1.6.2 for the form of the solution
to first-order kinetic reactions):

(1 − exp(−k t)) 1
(2.2)

An interesting thing happens if, for a given time and temperature, we plot
the value of Equation 2.2 given Equation 2.1 across activation energies, and
compare that to the approximation:

0, k t < 1
∆( k 1 , t) =  1 (2.3)
 1, k 1t ≥ 1

Figure 2.1 shows the result (actually plotting 1 − exp( −k 1t) and ∆( k 1 , t) for
a given temperature and time against activation energy). The center vertical
line is the approximation. The circles are the exponential functions at each
activation energy. The two outer vertical lines correspond to values of ν two
orders of magnitude off. The bounds shown by the outer lines are important
in the following discussion. Relationships 2.1 and 2.3 give the formal approx-
imation:
Ed = kT ln( νt) (2.4)

where Ed is the bounding value for activation energy, or the demarcation

energy. Reactions with activation energies less than Ed are essentially com-
plete for the given conditions of T, t, and ν.
Equation 2.4 is the thermal demarcation approximation. The fundamental
point of interest is that this approximation describes a relationship between all
values of Ea and all values of ν for given values of T and t. Essentially Equation
2.4 describes an approximate front between values of the pair (Ea, ν) corre-
sponding to reactions that are finished, and values of the pair corresponding

© 2004 by CRC Press, LLC

C4711_C02.fm Page 29 Saturday, February 28, 2004 10:29 AM

1.0
0.8
reaction extent
0.6
0.4
0.2
0.0

0.0 0.5 1.0 1.5 2.0

activation energy

FIGURE 2.1
Extent of reaction vs. activation energy plotted over the demarcation approximation.

to reactions that have not begun. In a plot of log(ν) (vertical) as a function

of Ea (horizontal), reactions represented by paired values of E and ν in the
region above and left of the line are complete and those below and right are
not. The term demarcation mapping derives from this combination of the
demarcation “line” (or more generally “curve” as shown below) on the plot
(“map”) of ν vs. Ea. Note that this description is now independent of the
need to consider only physical processes with distributions of activation
energies.
In the original use of the demarcation approximation as a method of scaling
time (Primak, 1955; Erdogan et al., 1994), the uncertainty was associated with
the distribution of activation energies within a material. Its validity was
driven by the law of large numbers, acting through the number of sites in
the material participating in macroscopic effects. The use of the demarcation
approximation promoted here (LuValle et al., 1998, 2000; LuValle, 2000) the
uncertainty now lies in the human lack of knowledge of the mechanism.
Certain bounds may apparently be used to ensure a conservative assessment
(LuValle, 2000) although these need either to be checked on a case-by-case
basis or to be understood from a general mathematical theory. Certainly, the
treatment of parameters here leans toward a Bayesian approach. At the end
of this chapter we follow this idea further, developing a generalized theory
for acceptance testing based on looking for processes shadowed by observed
degradation processes.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 30 Saturday, February 28, 2004 10:29 AM

With this notion of a demarcation approximation, we see that it would be

of great interest to the test engineer to be able to see how the demarcation
approximation for the time–temperature (possibly variable) exposure corre-
sponding to life compares to the demarcation approximation for a given
accelerated test. A good a priori accelerated test would be one that reached
the same set of (Ea, ν) pairs during the test, as are reached during life. Before
examining maps of such pairs over given stress exposures, we need to know
how to deal with Equation 2.4 when temperature is changed at a given time,
as happens both in life and in step stress accelerated testing. To see how
such a change is calculated, we note that at a given value of ν we can calculate
times and temperatures corresponding to the same value of Ea by simply
writing:

kT1 ln( νt1) = Ed = kT2 ln( νt2 ) (2.5)

or

T2

t1 =
( νt )
2
T1
(2.6)
ν

which implies that, after a step in temperature, Equation 2.4 can be re-
expressed:

  T1  

Ed = kT2 ln ν t2 +
( νt 1 ) T2
 (2.7)
  ν  
  


A key practical issue is how low the value of ν can be. As far as we have
found, there is no scientific reason to presume a lower bound. Diffusion
processes using this approximation can be expected to reach very low values.
However, for the one-step processes, quoted values typically range from 106
to 1018 Hz. Based on our experience with glasses and polymers, we have
chosen to use 10–5 Hz as a lower bound, but some careful thought needs to
be used. Even with this uncertainty, the maps provide significant information
on the relative efficacy of different accelerated tests. A reasonable, conserva-
tive alternative for thermal demarcation maps is to extend the maps to low
enough values of ν so that the corresponding Ed is 0 in all experiments and
life. In the example below this would be to 10–10.
To illustrate how demarcation mapping works, consider an operating life
consisting of a 4-h bake at 180°C, followed by 25 years at 50°C. We compare
this to an accelerated test with a 4-h bake at 180°C followed by 1 h at 260°C
in Figure 2.2. (Figure 2.2 and Figure 2.3 were produced using the code in
the complement to this chapter in Splus.) The shaded region (horizontal

© 2004 by CRC Press, LLC

C4711_C02.fm Page 31 Saturday, February 28, 2004 10:29 AM

20
1 2
1 2
1 2
1 2
1 2
15

1 2
1 2
1 2
1 2
1 2
10

1 2
1 2
1 2
log nu

12
12
5

12
12
12
12
12
0

12
12
12
12
12
-5

12
1
2
1
2
1
2
1
2
-10

1
2

0 1 2 3 4
Ea

FIGURE 2.2
Demarcation map of 180°C 4 h + 50°C 25 year life vs. 180°C 4 h + 260°C 1 h experiment.

lines) represents the region of reactions corresponding to life (it does not
include the burn-in as that occurs before the customer sees the device, and
failures in that part of the space will be eliminated by the burn-in). The space
between the lines labeled with 1s and that labeled with 2s represents the
region of reactions corresponding to the accelerated test. We note that the
accelerated test contains all reactions occurring during life only for values
of ν greater than 104.5 Hz. The vertical line is at 0 eV, representing the limit
of “reasonable” reactions. If we have reason to regard any reaction with
ν less than this to be an improbable contributor to failure, we can regard
survival of the accelerated test without failure as a basis for acceptance of
the product. If the value of ν might actually be close to 104 Hz, we may wish
to use the bounds on the approximation given by the outer lines in Figure
2.1 to guide us. Thus, we might wish to use an experiment such that the
accelerated test region contains all reactions occurring during life for values
of ν greater than 102 Hz. Figure 2.3 shows an experiment doing this. How-
ever, instead of having to extend the time on test by a factor of 100, we only
had to test at 260°C for 10 h, rather than 1 h.
A schema for an Excel spreadsheet is supplied as Table 2.1. The first six
rows are adjustable parameters; the remaining rows are formulae for columns.
We can also note for these diagrams that the accelerated tests include many
reactions not included during life. Thus, we see with this particular scheme
that there is a fair danger of seeing an artificial failure or degradation mode

© 2004 by CRC Press, LLC

C4711_C02.fm Page 32 Saturday, February 28, 2004 10:29 AM

20
1 2
1 2
1 2
1 2
1 2
15

1 2
1 2
1 2
1 2
1 2
10

1 2
1 2
1 2
log nu

1 2
1 2
1 2
5

1 2
1 2
1 2
1 2
0

12
12
12
12
12
-5

12
12
12
12
12
-10

12

0 1 2 3 4
Ea

FIGURE 2.3
Demarcation map of 180°C 4 h + 50°C 25 year life vs. 180°C 4 h + 260°C 10 h experiment.

during accelerated testing that would have nothing to do with service life
failure. The demarcation maps can be used to design tests with more specificity
as well, trading off time on test, with coverage, and the possibility of a false
alarm by using different time-stress trajectories. The maps can even be used to
design tests that peel back degradation or failure modes that analysis indicates
are artificial so we can look for failure mechanisms they may be masking. A
prior publication (LuValle, 1998) provides an example of this in a real problem.
Here we continue this example to show in principle how this could work.
Suppose that exposures of 4 h at 180°C and 10 h at 260°C result in several
failures, and that repeating the experiment with the second exposure at
300°C also results in failures. Suppose further that we have recorded the
times of these failures and find that they correspond to a process with an
activation energy of ~2.0 eV and a ν value of 1015 Hz. Because a failure mode
with these parameters will not occur during service life, the standard
assumption applied in accelerated testing is that we now have proved the
device reliable, because the only failure we have seen is one that will never
occur during life. However, the map points out that our accelerated tests
have not gotten close to a number of (Ea, ν) pairs that may be active during
life. The high-stress failure mode could be masking it in these tests. Thus,
what we want to do is to perform a test that minimally activates the (2.0,
1015) failure mode, but that emphasizes some of the lower values we have
missed. Such a test is described below.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 33 Saturday, February 28, 2004 10:29 AM

TABLE 2.1
Adjustable Terms Variable Name Example Value
Operating temperature OT 85
Operating time Otime (h) 24*365*25
Experimental temperature ET 260
Experimental time Etime (h) 1
Bake temperature BT 180
Bake time Btime (h) 4
Column Name/Symbol Column Formula Example Value
Log10(ν) lν 5
ν 10lν 100000

Demarcation energy (bake)

(8.625e − 5)(273 + BT) × 0.8239
ln(ν × Btime × 3600)

( 273+ BT )
Equivalent time under operating
conditions, tbO (ν × Btime × 3600) ( 273+ OT ) 1077.3
ν × 3600

( 273+ BT )
Equivalent time under experimental
conditions, tbE (ν × Btime × 3600) ( 273+ ET ) 0.1688
ν × 3600

(8.625e − 5)(273 + OT) ×

Demarcation energy (operating life) 0.9882
ln( ν × (Otime + t ) × 3600)
bO

(8.625e − 5)(273 + ET) ×

Demarcation energy (experiment) 0.9129
ln( ν × (Etime + t ) × 3600)
eO

The experiment below is possible if the device under question is very

small, so that its thermal inertia is very low (it can heat up and cool down
very quickly). This is possible with some microelectronic components, and
some optical fiber components. In Figure 2.4, life stress is as before. The
experimental stress is 180°C for 4 h, followed by 320°C for 15 s, followed by
170°C for 500 h. The shaded region again represents the one-step, first-order
reactions that can occur during life. The 320°C exposure is the region between
the line labeled 1 and the line labeled 2. It pulls the final exposure out close
to the failure mode regime (the F) but just short of it. The remaining exposure,
170°C for 500 h, provides significant aging of the device for most failure
mechanisms corresponding to the shaded regime, for values of ν down to
~10-6 Hertz. (where the activation energy is >0).
This exposure is illustrative of the potential of step-down stress tests to
explore regimes that might typically be masked in highly accelerated tests.
However, it is also illustrative of some of the cautions. In particular, this sort

© 2004 by CRC Press, LLC

C4711_C02.fm Page 34 Saturday, February 28, 2004 10:29 AM

20 1 2
3
1 2
3
1 2
3
1 2 3
1 23
15

1 23 F
1 23
1 23
1 23
1 23
10

12 3
12 3
12 3
log nu

12 3
12 3
5

1
2 3
1
2 3
1
2 3
1
2 3
1
2 3
0

1
2 3
1 3
2
1 3
2
1 3
2
1 3
2
-5

1 3
2
1 3
2
1 3
2
1 3
2
1 3
2
-10

1 3
2

0 1 2 3 4
Ea

FIGURE 2.4
Demarcation map of 180°C 4 h + 50°C 25 year life vs. 180°C 4 h + 320°C 15 s + 170°C 500 h
experiment.

of stress test would be impossible for a device with high thermal mass and
low surface area to volume. The map also points out that some potential
failure modes are economically untestable (very low ν values).
This section has illustrated some of the basic concepts of the demarcation
map, using the demarcation approximation derived from the Arrhenius rela-
tionship. In the next section, we look more closely at the particular problem
of developing an acceptance test for Arrhenius-driven failure modes.

2.2 Designing an Acceptance Test for a Purely

Thermal Process
If our goal is to design an acceptance test (a test such that if nothing fails
we will accept that the device is reliable), we would like the acceptance test
to provide coverage for all the processes we believe might occur during life,
while minimizing the chances of rejecting a good material system, with the
minimum cost and time expenditure. The plots and discussion above suggest
a single number that we can extract from the demarcation maps, which
provides a simple characterization of the degree of conservativeness of the
accelerated test with respect to accelerable failure modes. That number is

© 2004 by CRC Press, LLC

C4711_C02.fm Page 35 Saturday, February 28, 2004 10:29 AM

20 1
1
1
1
1
15

1
1
1
1
1
10

1
1
1
log nu

1
1
1
5

1
1
1
1
0

1
1
1
1
1
-5

1
1
1
1
1
-10

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Ea

FIGURE 2.5
Demarcation map of 40°C 30 year life vs. 90°C 100 h experiment.

the minimum value of ν where the accelerated test contains all the processes
occurring during life. This is the simplest use of demarcation mapping to
design accelerated tests that will arguably reach end of life. A conservative
value of ν for the purpose of including both processes running totally to
completion and some more complex processes (LuValle, 2000) is a value
three orders of magnitude lower than the lowest value for a simple single-
step chemical process in the materials under consideration.
Suppose the device to be life-tested is assembled using a low-melting-point
solder (e.g., indium–tin with a melting point of 110°C). We believe that the
device, which will see a maximum temperature of 40°C during manufactur-
ing and life, may have a thermally activated failure model. If we would like
to check whether the device could survive for 30 years, while testing at no
more than 90°C, how long will we have to test?
Constructing a demarcation map using the theory provided above, we can
construct a demarcation map for 40°C for 30 years (no bake) and plot over
it the demarcation map for 90°C 100 h.
The shaded region in Figure 2.5 contains the (Ea, ν) pairs that can be
expected to react during life, for those values of ν in the range of 10–10 to
1020. The region to the left and above the line marked with “1”s is the
corresponding region for the experiment. We see that the value of ν for which
the activation energies activated during the experiment are exactly those
activated during life is 1016. This is fairly high, as many common chemical

© 2004 by CRC Press, LLC

C4711_C02.fm Page 36 Saturday, February 28, 2004 10:29 AM

20 1
1
1
1
1
15

1
1
1
1
1
10

1
1
1
log nu

1
1
1
5

1
1
1
1
0

1
1
1
1
1
-5

1
1
1
1
1
-10

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Ea

FIGURE 2.6
Demarcation map of 40°C 30 year life vs. 90°C 500 h experiment.

reactions have ν values below 1016 Hz. A good moderately conservative value
would be 103 or lower.
Plainly, to have a more conservative test we need to lengthen the time at
90°C. If we change the time at 90°C to 500 h we obtain Figure 2.6, with a
crossover of approximately 1010.5. The reader can continue with trial and
error. Alternatively, the reader can complete the following exercises.

Exercise 2.2.1
Determine how to identify for fixed operating time and temperature and
experimental temperature the exact time needed to obtain equivalent
times at a given value of ν.

Exercise 2.2.2
Prove that for fixed temperature experiments, with the experiment at a
higher temperature than the operating temperature, this experiment will
be conservative for all values of ν larger than that found in 1.

Exercise 2.2.3
An alternative approach used in reliability engineering is to find the
smallest activation energy such that, for every activation energy above

© 2004 by CRC Press, LLC

C4711_C02.fm Page 37 Saturday, February 28, 2004 10:29 AM

that energy, the experiment is an end-of-life experiment. Prove that the

demarcation energy corresponding to the crossover value of ν is exactly
this activation energy.

2.3 Simple Temperature/Humidity Models

In the simplest form, the rate of reaction for combined temperature/humid-
ity models might be assumed to have the form:

 −E 
k 1 = νf ( RH , T) exp a  (2.8)
 kT 

where



 (RH)β [PL]

f ( RH , T) = (Vapor _ Pr essure)
β
[VP] (2.9)
 β
  RH 
   [BHT]
 1 − RH 

depending on the physics. The bottom expression was derived to model the
number of monolayers of gas adsorbed on a surface (Braunauer et al., 1938;
Klinger, 1991), and has been found to be useful in modeling reactions rates
occurring at material interfaces. The other two are used often, although
careful derivations of which situations are most appropriate for which model
are not available. As remarked in Complement 1.6.2, the first and second
representations are not statistically identifiable when coupled with the
Arrhenius relationship. Equation 2.8 results in a definition of demarcation
energy with the form:

( )
Ed = kT ln νf ( RH , T)t (2.10)

From this we can derive a similar result to that given in Equations 2.5, 2.6,
and 2.7, so that after a step in stress, the demarcation energy may be calcu-
lated:
  T1  

 
Ed = kT2 ln ν t2 +
( )
νf ( RH 1 , T1)t1 T2  
(2.11)
 
 νf ( RH 2 , T2 )  
  

© 2004 by CRC Press, LLC

C4711_C02.fm Page 38 Saturday, February 28, 2004 10:29 AM

2.4 Designing an Acceptance Test

for a Temperature/Humidity Model
Suppose that we start with a device that we suspect has a degradation mode
accelerated by both temperature and humidity (or that we want to argue
does not have any relevant failure mode so accelerated). Further suppose
that each device is given an 85°C, 85% relative humidity (RH), 100 h screen
by the customer, and then the device goes into a well-controlled telecom-
munications central office (maximum long-term temperature 45°C, maxi-
mum long-term humidity 42%.) for 20 years. How long and under what
conditions do we have to test to reach approximate end of life assuming a
minimal ν value of 103?
The freeware program provides facilities to construct an array of demar-
cation diagrams looking at each of these functions. For example, testing an
experiment run for 500 h at 90°C, 90% RH, with values of β 1 or 2 (reaction
rate proportional to concentration of water, or proportional to square of
concentration) results in the plots in Figure 2.7.
The lines without “+” show operating life (recall there is a burn-in defining
the beginning of operating life, hence two lines), while the lines with “+”
are experiments. The analytical solutions again are conservative although
finding each one becomes difficult. The partial vapor pressure in these plots
was calculated using the empirical relationship (Equation 1.3):

 0.4389 
VP = RH * exp 13.653 − .
 kT 

For the power law it is clear that we have not reached the goal proposed
above. The reader can input the software in the complement, and proceed
by trial and error until a solution is found, or the reader can perform the
following exercises:

Exercise 2.4.1
Determine how to identify for fixed humidity function operating time
and temperature, relative humidity, β, and experimental temperature and
relative humidity the exact time needed to obtain equivalent times at a
given value of ν.

Exercise 2.4.2
Prove that for fixed stress experiments, with the experiment at a higher
temperature than the operating temperature, this experiment will be
conservative for all larger values of ν found in 1. What happens for the
humidity functions?

© 2004 by CRC Press, LLC

 C4711_C02.fm Page 39 Saturday, February 28, 2004 10:29 AM
BHT Beta= 1 VP Beta= 1 PL Beta= 1

20

20

20
15

15

15
10

10

10
log nu

log nu

log nu
5

5
0

0
-5

-5

-5
-10

-10

-10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Ea Ea Ea

BHT Beta= 2 VP Beta= 2 PL Beta= 2

20

20

20
15

15

15
10

10

10
log nu

log nu

log nu
5

5
0

0
-5

-5

-5
-10

-10

-10
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Ea Ea Ea

FIGURE 2.7
Array of temperature/humidity demarcation maps of life consisting of 85°C 85% RH screen followed by 20
years at 40°C 42% RH vs. an experiment of 90°C 90% RH for 500 h.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 40 Saturday, February 28, 2004 10:29 AM

Exercise 2.4.3
What advantages do the plots have over the crossover points in terms
of information about artificial and uncovered failure mechanisms in ac-
celerated tests?

Exercise 2.4.4
Check the crossover point for the vapor pressure and power law curves.
Which is more conservative? Why?

2.5 Mechanical Cycling Models

A simple empirical model of the effect of cyclic mechanical strain is the
Coffin–Manson model (Manson, 1981). The model has the form:

ε p = MN zf (2.12)

where εp is plastic strain, M and z are material-related constants, and Nf is the

number of cycles the device withstands until failure. If we do not reach failure,
Equation 2.12 provides a way of judging equivalent amounts of internal dam-
age to the material (degradation). Miner’s rule (Miner, 1945) supports this use
of the Coffin–Manson equation. Empirically for notched metal samples, z is
close to –½, although other values have been reported. Assuming that
εp corresponds to a self-diffusion length for the material on each cycle (propor-
tional to the nonrecoverable plastic strain), then z close to –½ corresponds to
failure occurring when diffusion has occurred to a certain constant extent.
Ignoring for the moment the diffusion argument, which would fix z, then
Equation 2.12 can be transformed to

1=
( M N ) ⇔  1 =  1  log( M
1/z
f 1/z
)
   Nf (2.13)
 z   log( ε ) 
(ε )
1
z p
p

The equation on the left of Equation 2.13 corresponds to Equation 2.4. From
this the step stress criterion to go from plastic strain ε p1 to ε p2 is

( )
log ε p 2

 1z 
 M Nf 
( )
log ε p 1

 
N f 2* = 1
(2.14)
z
M

© 2004 by CRC Press, LLC

C4711_C02.fm Page 41 Saturday, February 28, 2004 10:29 AM

Miner’s rule (Miner, 1945) gives empirical support to the use of Equations
2.13 and 2.14 to construct demarcation maps with varying amounts of plastic
( ) (
strain. Our implementation of a demarcation map plots log10 M 1 Z vs. − 1 z . )
In this way, the plot can be interpreted the same way as the thermal demar-
cation map, with values above and to the right of the drawn line values that
will occur during the stress trajectory.

[ ( )(
The maps are drawn in the log10 M 1 z , − 1 z )] plane. However, the value
of z = −1 2 is marked with a red line to aid in interpreting the plot in line
with the diffusion hypothesis. Understanding other values of z would
require additional theoretical analysis.

2.6 Acceptance Testing for Mechanical Cycling Induced

by Thermal Cycling
Typically, in electronic equipment, mechanical cycling is done indirectly,
either through thermal cycling, vibration, or power cycling of the devices.
Thus, to employ this calculation directly some side experiments must be
done to ascertain the relationship between the applied cycle and the plastic
strain seen in the device under question. Further, in chemistry, values of ν
are sometimes estimated experimentally, and there is some theory (Glasstone
et al., 1941) relevant to the task. It is not clear that values of M or M 1 z are
similarly attainable anywhere. The main assumption in what follows is that
the primary effect of the thermal cycle is that the mismatch between the
coefficients of thermal expansion for different materials drive a mechanical
cycling in a material seeing significant plastic strain. The chemical effects
due to raised temperature are assumed negligible.
Suppose that we have a problem where a device might be expected to be
turned on once every hour for 2 years in normal operation, and the plastic
strain applied to the component of interest is 1e–4 each time it is turned on.
To ensure that very bad devices are not shipped, the device might undergo
ten thermal cycles corresponding to a plastic strain of 1e–3. Following a
standard rule-of-thumb engineering practice for developing such environ-
mental stress screens that we have encountered, suppose that 20 devices
were subjected to 100 thermal cycles, and none failed.
The question a conservative reliability engineer would ask is, were the 100
thermal cycles enough to make sure we have not increased the failure rate
by applying this test? As a first cut we can specify values of M 1 z to be used
in the plot of 10 −5 − 10 10 .

© 2004 by CRC Press, LLC

C4711_C02.fm Page 42 Saturday, February 28, 2004 10:29 AM

10^-2 10^0 10^2 10^4 10^6 10^8 10^10

M^(1/z)
10^-5

0 1 2 3 4
-1/z

FIGURE 2.8
Coffin–Manson demarcation maps of life consisting of 10-cycle screen at 1e–3 plastic strain
followed by 2 years of hourly cycles at 1e–4 plastic strain vs. an experiment of a 10-cycle screen
at 1e–3 plastic strain followed by 100-cycle “safety test” at 1e–3 plastic strain.

The map, given in Figure 2.8, shows that even if we consider only the
( )
value of − 1 z corresponding to diffusion, the “safety experiment” of 100
thermal cycles has not tested to end of life, so does not indicate if we will
see any serious increase in failure due to this testing or not. The line without
“+” is the line representing the screen followed by life, while the lower line
with “+” is the screen followed by the safety test; the top line is the screen.
The vertical line is the value of − 1 z = 2 . ( )
As two alternatives, we can consider thermal cycles with identical strain
as the screen, and thermal cycles at a lower strain level than the screen.
Figure 2.9 runs the test to 170 cycles. Now the safety test at 1e–3 plastic strain
does reach end of life under the diffusion hypothesis, but not for more
conservative values of –1/z. Figure 2.10 shows what happens if we drop the
plastic strain in the safety test to 3e–4 for 500, 2500, and 5000 cycles (respec-
tively, the second, third, and fourth lines with “+” from the top).

Exercise 2.6.1
Determine algebraically how many cycles are required in the safety ex-
periment to ensure no failures by end of life assuming a – 1/z value of
2, and of 1, for a plastic strain value of 3e–4.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 43 Saturday, February 28, 2004 10:29 AM

10^-2 10^0 10^2 10^4 10^6 10^8 10^10

M^(1/z)
10^-5

0 1 2 3 4
-1/z

FIGURE 2.9
Coffin–Manson demarcation maps of life consisting of 10-cycle screen at 1e–3 plastic strain
followed by 2 years of hourly cycles at 1e–4 plastic strain vs. an experiment of a 10-cycle screen
at 1e–3 plastic strain followed by 170-cycle “safety test” at 1e–3 plastic strain.
10^-2 10^0 10^2 10^4 10^6 10^8 10^10
M^(1/z)
10^-5

0 1 2 3 4
-1/z

FIGURE 2.10
Coffin–Manson demarcation maps of life consisting of 10-cycle screen at 1e–3 plastic strain
followed by 2 years of hourly cycles at 1e–4 plastic strain vs. an experiment of a 10-cycle screen
at 1e–3 plastic strain followed by 3e–4 plastic strain at 500, 2500, and 5000 cycles.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 44 Saturday, February 28, 2004 10:29 AM

2.7 Computational Demarcation Mapping

It is possible to use the simple temperature demarcation map to bound more
complex kinetic processes (LuValle, 2000), but that is only proved for con-
stant stress situations for a small set of models (the two-arrow kinetic models
that can be accelerated). An alternative to the use of elegant approximations
such as those contained in Equations 2.4 through 2.7 is to perform a numer-
ical calculation, over a large array of parameters and identify if the extent
of reaction of interest is larger from an accelerated experiment than from life.
For an example we consider a problem that appears in devices made from
metal films deposited at high temperatures on substrates with lower coeffi-
cients of thermal expansion. The problem is stress voiding, and is often
related to electromigration. Stress voiding is an odd kinetic process, in that
the “effective activation energy” (the apparent activation energy obtained
by plotting temperature vs. time to failure) has a curvature with a maximum.
The model we use is a simple version of the creep process that Krauss and
Erying (1975) describe combined with a compartment model. The model of
Krauss and Erying is a model in which microscopic movement involves
overcoming an activation barrier, and stress preferentially changes the acti-
vation energy. In particular the model has the form:
k k k
f → f → f →
A1 A2 ... A (2.15)
←
k
 ←
k
 ← k
 n
b b b

where

(
 E − stress.coef × T − T ) 
( )
pow a 0 +
k f = fluence ν exp  − (2.16)
 kT 
 

and

(
 E + stress.coef × T − T ) 
( )
pow a 0 +
kb = fluence ν exp − (2.17)
 kT 
 

T0 is the temperature where the film and the underlying material are at 0
stress, the notation ( z) denotes we are setting the value as z if it is positive,
+
and 0 if it is negative. We could take an arbitrarily complex kinetic model if
we wished, but this is sufficiently complex. This model could be used as a
combined model of electromigration and stress voiding, with fluence denot-
ing current density. We assume there is no fluence effect. Then in the simplest
version of the real problem, there are four unknown parameters to cause us
( )
concern: ν, Ea , stress.coef , T0 , and there is a moderately complex relation-
ship based on Equations 2.15 through 2.17.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 45 Saturday, February 28, 2004 10:29 AM

Strictly, this sort of model is a local microscopic model of one side of a

region where the void is forming. Hence depletion of the early states corre-
sponds to formation of the void. Thus, a good surrogate for the actual
degradation is the concentration of material left in the first few states.
To evaluate this model, we need to proceed in four steps:

1. Write a program to evaluate the degradation model, including both

kinetic equations, and its translation to an observable at given
parameter values for given stress trajectories. Our approach uses a
finite set of compartments.
2. Write a program to step through the parameters, evaluating a stress
trajectory, and recording the observable.
3. Compare observables for life and experiment trajectories.
4. Iterate until a combination of experiments is decided.

Steps 1, 2, and 3, can be done in environments such as Mathcad®, Mathe-

matica®, and Matlab®. For demonstration purposes, the kinetic modeling
software described in the complement to Chapter 3, written in Splus is used.
The model we use in studying this problem is of the form from Equations
2.15 through 2.17 with 20 states. Total depletion of the first five states is taken
as the kinetic marker for degradation or failure.
We assume for this device that we are concerned only with the time the
device is actually on the shelf, and that when it is in use it runs hot enough
so that there is 0 stress. We assume that the shelf life is 5 years. The parameter
values we actually do the mapping for are as follows:

() (
log10 ν = −4, 0, 4, 8, 12, 16, 20 )
E = (0.3, 0.6, 0.9, 1.2, 1.5)
a

T = (200, 350, 500), or 600

0

stress.coef = (1e − 6, 1e − 5, 1e − 4, 1e − 3, 1e − 2)

T0 varies through 200, 350, and 500 for the values in life. In six experiments
it follows this variation, and in the remaining experiments it is set to 600.
This allows us to examine the effect of T0 both if it is uncontrolled (but
constant through manufacture), and if it can be used as an experimental
variable for acceleration.
Figure 2.11 is a demarcation map comparing storage life at 50°C for 5 years
to an accelerated experiment at 150°C for 5000 h. Each graph in the plot
corresponds to the values of pow (the exponent of the fluence term = 0
identically), T0, and stress.coef being fixed to the values shown above that
graph for the life stress. In this case the T0 values changed in the experiment
in a corresponding fashion. The horizontal axis of each graph is the activation

© 2004 by CRC Press, LLC

C4711_C02.fm Page 46 Saturday, February 28, 2004 10:29 AM

TABLE 2.2
Evaluation of Potential Experiments
Number
Temperature Time T0 (of 525 possible)
100 5000 h 200,350,500 372
150 5000 h 200,350,500 363
100 2 years 200,350,500 357
150 2 years 200,350,500 350
200 2 years 200,350,500 405
300 2 years 200,350,500 413
50 2 years 600 196
100 2 years 600 194
150 2 years 600 195
200 2 years 600 269
100* 1000 h 600 218
250* 1000 h 600 270
Combo (two above*) 190

energy (Ea) in electron volts, and the vertical axis is log 10 ( ν) . An L in a spot
means that life will result in more depletion in the first five compartments
than the experiment, so the experiment cannot be thought of as having
reached end of life. An E implies that the experiment will result in more
depletion, implying that for the parameter values corresponding to that
position we have reached beyond the end of life degradation in the experi-
ment. (A key assumption here is that end of life is controlled only by the
function of the state vector we are examining, and the environment influ-
ences end of life only through that state vector).
A very simple crude measure of the efficiency of an experiment is the
number of points in the map where life dominates experiment. For a fixed
map, such as here, the fewer points dominated by life, the more likely the
experiment is to identify regions where failure can occur. Although this
actually must be weighed by the likelihood of each parameter point, for the
interest of simplicity we start with the simple measure and allow Chapter 4
to consider more sophisticated measures. Table 2.2 shows the experiments
tried vs. the number of points where life dominates.
There are several points of interest that can be derived from this table.
First, if T0 can be controlled, it makes a far more effective accelerant than
temperature for probing for this phenomenon. Second, increasing T0 shifts
the optimal temperature for achieving high coverage. Finally, combining
experiments cleverly can result in significant shortening of the time required
to attain a certain level of coverage.
The particular coverage by the 2-year experiment at 100°C with a T0 of
600°C is shown in Figure 2.12 and can be contrasted to the coverage shown
in Figure 2.13 for the combined experiment run at 1000 h each. The differ-
ences are actually slight, although they do exist.

© 2004 by CRC Press, LLC

 C4711_C02.fm Page 47 Saturday, February 28, 2004 10:29 AM
0 200 1e-006 0 200 1e-005 0 200 0.0001 0 200 0.001 0 200 0.01

20

20

20

20

20
L L L L L L L L L L L L L L L L L L L L L L L L L

L L L L E L L L L E L L L L E L L L L L L L L L L

15

15

15

15

15
L L L L E L L L L E L L L L E L L L L E L L L L L

10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
L L L E E L L L E E L L L E E L L L E E L L L L L

5
L L E E E L L E E E L L E E E L L E E E L L L L L
0

0
L E E E E L E E E E L E E E E L L E E E L L L L L

L E E E L L E E E L L E E E E L L E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 350 1e-006 0 350 1e-005 0 350 0.0001 0 350 0.001 0 350 0.01
20

20

20

20

20
L L L L L L L L L L L L L L L L L L L L L L L L L

L L L L E L L L L E L L L L E L L L L L L L L L L
15

15

15

15

15
L L L L E L L L L E L L L L E L L L L E L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
L L L E E L L L E E L L L E E L L L E E L L L L L
5

5
L L E E E L L E E E L L E E E L L L E E L L L L L
0

0
L E E E E L E E E E L E E E E L L E E E L L L L L

L E E E L L E E E L L E E E E L L E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 500 1e-006 0 500 1e-005 0 500 0.0001 0 500 0.001 0 500 0.01
20

20

20

20

20
L L L L L L L L L L L L L L L L L L L L L L L L L

L L L L E L L L L E L L L L E L L L L L L L L L L
15

15

15

15

15
L L L E E L L L E E L L L L E L L L L L L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
L L L E E L L L E E L L L E E L L L L E L L L L L
5

5
L L E E E L L E E E L L E E E L L L E E L L L L L
0

0
L E E E E L E E E E L E E E E L L L E E L L L L L

L E E E L L E E E L L E E E E L L L E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

FIGURE 2.11
Computational demarcation map of stress voiding, comparing life at 50°C for 5 years vs. an experiment
at 150°C for 5000 h. L means that the reaction extent for life dominates the reaction extent for the experiment
at that combination of parameters. E means the reverse.

© 2004 by CRC Press, LLC

 C4711_C02.fm Page 48 Saturday, February 28, 2004 10:29 AM
0 200 1e-006 0 200 1e-005 0 200 0.0001 0 200 0.001 0 200 0.01

20

20

20

20

20
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E E E E E E E E E E E L L L L L

15

15

15

15

15
E E E E E E E E E E E E E E E E E E E E L L L L L

10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
E E E E E E E E E E E E E E E E E E E E L L L L L

5
0 E E E E E E E E E E E E E E E E E E E E L L L L L

0
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E L E E E E L E E E E L E E E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 350 1e-006 0 350 1e-005 0 350 0.0001 0 350 0.001 0 350 0.01
20

20

20

20

20
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E E E E E E E E E E E L L L L L
15

15

15

15

15
E E E E E E E E E E E E E E E E E E E E L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
E E E E E E E E E E E E E E E E E E E E L L L L L
5

5
E E E E E E E E E E E E E E E E E E E E L L L L L
0

0
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E L E E E E L E E E E L E E E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 500 1e-006 0 500 1e-005 0 500 0.0001 0 500 0.001 0 500 0.01
20

20

20

20

20
L L L L L L L L L L L L L L L L L L L L L L L L L

L L L L E L L L L E L L L L E L L L L L L L L L L
15

15

15

15

15
L L L E E L L L E E L L L L E L L L L L L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
L L L E E L L L E E L L L E E L L L L E L L L L L
5

5
L L E E E L L E E E L L E E E L L L E E L L L L L
0

0
L E E E E L E E E E L E E E E L L E E E L L L L L

E E E E L E E E E L E E E E L L E E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

FIGURE 2.12
Computational demarcation map of stress voiding, comparing life at 50°C for 5 years vs. an experiment at
100°C for 2 years, with T0 = 600°C.

© 2004 by CRC Press, LLC

 C4711_C02.fm Page 49 Saturday, February 28, 2004 10:29 AM
0 200 1e-006 0 200 1e-005 0 200 0.0001 0 200 0.001 0 200 0.01

20

20

20

20

20
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E E E E E E E E E E E L L L L L

15

15

15

15

15
E E E E E E E E E E E E E E E E E E E E L L L L L

10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
E E E E E E E E E E E E E E E E E E E E L L L L L

5
E E E E E E E E E E E E E E E E E E E E L L L L L
0

0
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E E E E E E E E E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 350 1e-006 0 350 1e-005 0 350 0.0001 0 350 0.001 0 350 0.01
20

20

20

20

20
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E E E E E E E E E E E L L L L L
15

15

15

15

15
E E E E E E E E E E E E E E E E E E E E L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
E E E E E E E E E E E E E E E E E E E E L L L L L
5

5
E E E E E E E E E E E E E E E E E E E E L L L L L
0

0
E E E E E E E E E E E E E E E E E E E E L L L L L

E E E E E E E E E E L E E E E E E E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

0 500 1e-006 0 500 1e-005 0 500 0.0001 0 500 0.001 0 500 0.01
20

20

20

20

20
L L L L L L L L L L L L L L L L L L L L L L L L L

L L L L E L L L L E L L L L E L L L L L L L L L L
15

15

15

15

15
L L L E E L L L E E L L L L E L L L L L L L L L L
10

10

10

10

10
Diff0

Diff0

Diff0

Diff0

Diff0
L L L E E L L L E E L L L E E L L L L E L L L L L
5

5
L L E E E L L E E E L L E E E L L L E E L L L L L
0

0
L E E E E L E E E E L E E E E L L E E E L L L L L

L E E E E L E E E E L E E E E L L E E E L L L L L
0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4 0.4 0.6 0.8 1.0 1.2 1.4
Ea Ea Ea Ea Ea

FIGURE 2.13
Computational demarcation map of stress voiding, comparing life at 50°C for 5 years vs. combined experiment
100°C for 1000 h, and another experiment of 250°C for 1000 h, both with T0 = 600°C.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 50 Saturday, February 28, 2004 10:29 AM

2.8 Beta Binomial Interpretation of 0 Failures

Consider Figure 2.12. This figure represents an experiment where some
number of devices were aged for 2 years at 100°C. For each point with an E
in the diagram, if the devices obey an aging law corresponding to that point,
then the experiment aged each of those devices beyond the level where they
would be aged during life. Thus, for no failures during the experiment, at
those points where an E sits in the diagram, we can assume that the exper-
iment provides an upper bound for failure during life. From Complement
( )
1.6.4, if we start with a β α , β prior distribution for each point, then at each
point with an E, our posterior distribution would be a β α , β + n . Thus, at ( )
each point with an E, the odds we would consider a fair bet for failure have
changed from α : β (corresponding to an α α + β probability of failure) to
α : β + n (corresponding to an α α + β + n probability of failure). On the other
hand, at each point where an L occurs, there has been less aging than would
occur during life. Assuming that Murphy’s law rules the universe, then the
monotone function relating failure and degradation to the state vector could
be sharply concave upward, so the noticeable degradation or failures could
be waiting until toward the end. Thus, where an L is written, we have no
information on bounding the probability of failure during life.
Ultimately, the purposes of our accelerated tests are to inform the people
in the business/operation area how to make a decision. Thus, these diagrams
must be converted somehow to a number that gives us the ability to decide
between a yes and a no. A good number to base that on is the probability
of failure during life. To reach that in a consistent way, we have to extend
the Bayesian paradigm (Ferguson, 1967; DeGroot, 1970) that we have been
using. We have to assign a prior probability to each point in the hyper-
rectangle in Figure 2.11. Then before the experiment, assuming no failures,
the probability of failure is

Γ (α + β)
1
α α
∫ ∫ P Γ (α)Γ (β) P α −1
(1 − P) β−1
dPdπ( φ) =
α +β ∫ dπ(φ) = α + β (2.18)
H 0 H

The term π( φ) is simply the probability distribution over the hyper-rectangle

H depicted in Figure 2.11; φ is the point representing a particular set of the
parameters. Letting E denote the region of H where the experiment domi-
nates life and letting L denote the region where life dominates the experi-
ment, by our argument above, the posterior probability that a failure or
noticeable degradation will occur during life after an experiment has
occurred is

© 2004 by CRC Press, LLC

C4711_C02.fm Page 51 Saturday, February 28, 2004 10:29 AM

1
( ) P 1− P
Γ α +β
∫ ∫ Γ (α)Γ (β + n) ( ) ()
β + n− 1
α −1
P dPdπ θ
E 0

Γ (α + β)
1

∫ ∫ Γ (α)Γ (β) P (1 − P) ()
β−1
α −1
+ P dPdπ θ
L 0 (2.19)
α α
=
α +β+n ∫ dπ(θ) + α + β ∫ dπ(θ)
E L

=
α
α +β+n
×πE +
α
α +β
()
× 1− π E ( ( ))
From the final portion of Equation 2.19 we see that the best experiment
(disregarding costs) is the one with the highest prior probability on the points
where the experiment will dominate life. Thus, the best experiment to reduce
the risk associated with assuming no failures will occur during life is the
experiment most likely to disprove the hypothesis that no failures will occur
during life. This result is important enough to emphasize it.

2.9 An Extrapolation Theorem

Assume the observable, y, is Bernoulli with probability p( φ, Ei ) . Assume the
following:

1. That p( φ, Ei ) is unknown, but that there is a known ordering

induced by the experiments. Thus, if Ei ≠ E j , then knowing φ implies
( )
t h a t w e k n o w i f p( φ , Ei ) ≤ p φ , E j , p( φ , Ei ) = p φ , E j , o r( )
( )
p( φ , Ei ) ≥ p φ , E j . This assumption is the key assumption tying the sta-
tistical theory developed here to the substantive science. In the situation
described in this chapter, the theory of computational demarcation maps is
what provides us with the ability to make this statement.
2. That there are desirable and undesirable outcomes (e.g., survival
and failure), and the undesirable outcomes correspond to y = 1.
3. That for each experiment Ei there is a corresponding sample size
ni of the Bernoulli y.
( )
4. That Lθ ( φ) = nρ p φ , Eρ . Here Lθ ( φ) is the loss if φ is true, and we
extrapolate using θ. Eρ is the special experiment corresponding to
what we are trying to extrapolate to. Define the risk of an experiment
R(Ei ) as the expected loss given the prior on the φ.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 52 Saturday, February 28, 2004 10:29 AM

5. That costs for all Ei (except Eρ ) are constant.

6. That p(θ, Ei ) ≡ 0 for all i.

Assume that at each point φ a beta prior is specified for p( φ, Ei ) and assume
that it is the same beta prior at all φ.

{ (
Define Sij = φ p( φ , Ei ) ≥ p φ , E j )} , and Θ θ as the set of all the φ in the neigh-
borhood of θ that we wish to consider. Define an estimate determined from
( )
Ei for p φ, E j to be a strongly determined upper estimate only if φ ∈Sij .
Then we have:

LEMMA
Under the assumptions, and assuming ni ≡ n , the experiment that, with a
null result (no observable degradation or failure), minimizes risk using
strongly determined upper estimates is that experiment for which π Siρ is ( )
maximized.

PROOF
1

( )
R(Ei ) = E Lθ ( φ) Ei = nρ
∫ ∫ dpp(φ, E )β(α , β E , φ)dπ(φ)
Θθ 0
ρ i
(2.20)

Here the second equality follows from assumption 4, and β α , β Ei , φ is the ( )

beta posterior density given experiment Ei and φ. Using the strong determi-
nation criteria, this is a β(α , β + n) density over Siρ , and a β(α , β) density
over Θθ \Siρ . Thus, under strong determination we can decompose Expres-
sion (2.20) into

 
 α   α 
( )
R Ei = nρ   
  α + β + n
d
∫ ()
π φ +  
 α + β
d
∫ ()
π φ 

 Siρ Θθ \Siρ  (2.21)
   α  
= nρ  
α
( )
 π Siρ + 
  α + β + n
( ( ))
 1 − π Siρ 
 α + β 

Clearly, Expression 2.21 is minimized by maximizing π Siρ . Note this is ( )

exactly the described result. In particular, Siρ is precisely the set of points
where no signature events during experiment i strongly implies an upper
bound on the probability of a signature event during Eρ .

© 2004 by CRC Press, LLC

C4711_C02.fm Page 53 Saturday, February 28, 2004 10:29 AM

2.10 Summary
In this chapter a radically different approach to the extrapolation of acceler-
ated tests has been described. This theory is based on the following notion:
The information on reliability (for ultrahigh-reliability devices) in an accelerated
test is in the failure and degradation free periods prior to any observable degradation
or failure. We have provided a theoretical basis for using this approach assum-
ing that tests can be constructed in which no failures are expected to occur
given the existing models of failure, and assuming that the failure modes
are described by a rather simple one-step kinds of kinetic models.
The complement to this chapter provides instructions for using the down-
loadable computer code in Splus for analytical demarcation mapping.
In Chapter 4, we extend the notion of demarcation mapping to considering
more complex (but still first-order) chemical kinetic processes as possible
alternatives to whatever model currently fits the failure and degradation
data obtained.

2.11 Complements to Chapter 2

The purpose of this section is to provide examples of using the software for
analytical demarcation maps. Examples of use of the software for computa-
tional demarcation maps must wait until the complement for Chapter 3,
when the kinetic modeling software is provided. Complement 2.11.1 contains
an extension of the extrapolation theorem in Sections 2.8 and 2.9 to multiple
experiments.

2.11.1 Demarcation Maps for Multiple Experiments

In most real problems we will wish to conduct an array of experiments. The
lemma in Section 2.9 only deals with choosing a single experiment. The
calculation of posterior probabilities for multiple experiments is relatively
straightforward. In particular, the strong determination criterion defines a
partition on the space of parameters into Siρ and Θθ \Siρ for each experi-
ment i. For n experiments this gives us partitions:

 S1ρ = S11ρ Θθ \ S1ρ = S10ρ 

 1 
 S1ρ S20ρ  (2.22)
 ... ... 
 1 0 
 S1ρ Smρ 

© 2004 by CRC Press, LLC

C4711_C02.fm Page 54 Saturday, February 28, 2004 10:29 AM

A complete partition of Θθ into mutually non-intersecting sets can be con-

structed from this. In particular define:

Pk = IS i
binary( k , i )
iρ

where binary( k , i) is the ith digit in the binary representation of k, k from 1

to 2 m . The posterior probability for the m NULL experiments, the ith exper-
iment with sample size ni , from which risk can be calculated is

expected posterior probability of failure

 
 
α × π Pk ( ) (2.23)

= ∑ m





k
α +β + ∑(
i= 1
binary( k , i) × ni 

)

With an array of fixed environmental settings, the problem of optimization

becomes a problem of distributing samples between the different experi-
ments. For simplicity assume a fixed total number N of experimental devices,
then Equation 2.23 can be re expressed:

expected posterior probability of failure

 
 
α × π Pk ( ) (2.24)

= ∑ m





k
α +β+
i =1
∑(
binary( k , i) × Np i 

)

where p i is a proportion, and

1= ∑p
i =1
i

Ignoring for the moment that devices come in whole numbers, we can
apply elementary calculus. Define from the constraint

m −1

pm = 1 − ∑p i =1
i

© 2004 by CRC Press, LLC

C4711_C02.fm Page 55 Saturday, February 28, 2004 10:29 AM

Then the partial derivatives take the form:

∂
∂p i
(
expected posterior probability of failure = )
  
  
 binary( m , i) × α × π Pk ( )  

 k∑  m  
2
 


 α + β +

∑( binary( k , i) × Np i  )
 


 i =1

N 
  
  
  binary( k , i) × α × π Pk( ) 
 −
 
∑  m  
2


∑( )
k
   α + β + binary ( k , i) × Np i 

     
 i =1

We leave it as an exercise to the interested reader to show that the common

0 point for this derivative is a minimum given the constraints on the p i .

2.11.2 Using the Freeware

The purpose of this section is to acquaint the user with the code associated
with this section. We make primitive use of the native GUIs available in
Splus in this section, to provide simple drop-down menus and popup com-
mands. The Appendix to the book describes how to install the freeware.
Once all the code is available in Splus, the simplest way to use it is using the
drop-down menu and the primitive GUIs native to Splus. For those who have
used SPLIDA™ coded by Bill Meeker, the interface here will seem primitive.
We encourage ambitious readers to modify the source code to their conve-
nience. However, it would be wise to retain the original as a backup.
If the reader has followed the instructions in the appendix, there should
be an entry in the top toolbar “Demarcation maps.” Left-clicking the mouse
button on this entry gives a drop-down menu with the list

Life-emphasized temperature demarcation map

Temperature/humidity demarcation map
Coffin Manson demarcation map
Create function for direct kinetic evaluation using kinetic model
Plot extent matrices from direct kinetic eval
Plot INT of extent matrices from direct kinetic eval

© 2004 by CRC Press, LLC

C4711_C02.fm Page 56 Saturday, February 28, 2004 10:29 AM

FIGURE 2.14
GUI for thermal demarcation map function in Splus.

Slot for name to store numerical

output of function
Time vector for experiment

Temperature
vector for life
(°C)
Title of plot
(in quotes)

Time vector for

life (hours) Vector of
log 10 (ν) value
Maximum activation
Temperature vector for energy for plot
experiment (°C)

FIGURE 2.15
Annotated GUI for thermal demarcation map function in Splus, filled in to produce Figure 2.4.

The last three commands are connected to evaluating computational demarca-

tion maps, so we will wait until the complement to Chapter 3 to see how to
use them. The first three items are the three analytical demarcation maps. The
simplest way to check them is against the calculations presented in the book.

2.11.2.1 Thermal Demarcation Maps

To access the GUI for simple analytical thermal demarcation maps, left-click
on “Demarcation maps” above the toolbar in Splus, and then left-click on
the “Life emphasized temperature demarcation map” entry in the demarca-
tion map drop-down menu. This brings up the default GUI for the function
“demarc.Ea.emph1,” shown in Figure 2.14. In Figure 2.15 we show this GUI
filled out to construct Figure 2.4 in the text, and provide a guide to each slot.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 57 Saturday, February 28, 2004 10:29 AM

As configured, left-clicking with the mouse on the apply button causes the
function to execute, producing Figure 2.4 and bringing up the “report”
window in Splus, where the numerical output is dumped. The numerical
output is a list, printed below. Note that vectors in the input use the usual
Splus notation beginning with a c(…).

$order.1.life:
matx mat.ret vec.ret
[1,] -5 -0.07571782 0.2499654
[2,] -4 0.01424706 0.3141129
[3,] -3 0.10421194 0.3782609
[4,] -2 0.19417682 0.4424103
[5,] -1 0.28414170 0.5065632
[6,] 0 0.37410657 0.5707249
[7,] 1 0.46407145 0.6349086
[8,] 2 0.55403633 0.6991479
[9,] 3 0.64400121 0.7635266
[10,] 4 0.73396608 0.8282502
[11,] 5 0.82393096 0.8938027
[12,] 6 0.91389584 0.9612132
[13,] 7 1.00386072 1.0322211
[14,] 8 1.09382560 1.1086037
[15,] 9 1.18379047 1.1906019
[16,] 10 1.27375535 1.2766537
[17,] 11 1.36372023 1.3649038
[18,] 12 1.45368511 1.4541597
[19,] 13 1.54364999 1.5438388
[20,] 14 1.63361486 1.6336898
[21,] 15 1.72357974 1.7236094
[22,] 16 1.81354462 1.8135564
[23,] 17 1.90350950 1.9035141
[24,] 18 1.99347437 1.9934762
[25,] 19 2.08343925 2.0834400
[26,] 20 2.17340413 2.1734044

$order.1.expt:
matx mat.ret vec.ret vec.ret
[1,] -5 -0.07571782 -0.07568411 0.1107292
[2,] -4 0.01424706 0.01430510 0.1987238
[3,] -3 0.10421194 0.10431185 0.2867194
[4,] -2 0.19417682 0.19434876 0.3747162
[5,] -1 0.28414170 0.28443745 0.4627144
[6,] 0 0.37410657 0.37461487 0.5507146
[7,] 1 0.46407145 0.46494374 0.6387176
[8,] 2 0.55403633 0.55552947 0.7267251
[9,] 3 0.64400121 0.64654624 0.8147398
[10,] 4 0.73396608 0.73827349 0.9027669
[11,] 5 0.82393096 0.83113808 0.9908161
[12,] 6 0.91389584 0.92574114 1.0789056
[13,] 7 1.00386072 1.02282347 1.1670707
[14,] 8 1.09382560 1.12311511 1.2553813
[15,] 9 1.18379047 1.22708066 1.3439780
[16,] 10 1.27375535 1.33470647 1.4331434
[17,] 11 1.36372023 1.44550735 1.5234262

© 2004 by CRC Press, LLC

C4711_C02.fm Page 58 Saturday, February 28, 2004 10:29 AM

[18,] 12 1.45368511 1.55874689 1.6158006

[19,] 13 1.54364999 1.67368908 1.7116992
[20,] 14 1.63361486 1.78974188 1.8125674
[21,] 15 1.72357974 1.90648713 1.9189315
[22,] 16 1.81354462 2.02365199 2.0299552
[23,] 17 1.90350950 2.14106672 2.1441137
[24,] 18 1.99347437 2.25862872 2.2600638
[25,] 19 2.08343925 2.37627697 2.3769438
[26,] 20 2.17340413 2.49397556 2.4942834

$crossover:
[1] 0.8400869 0.4496849

The component order.1.life is the matrix of demarcation points correspond-

ing to life. The first column is log 10 ( ν) , the second is the demarcation energy
corresponding to that value of ν after 4 h at 180°C. The third is the demar-
cation energy after a subsequent 25 years at 50°C.
The component order.1.expt is the similar matrix for the experiment tem-
perature trajectory. The component labeled crossover is the estimated values
of Ea and log 10 ( ν) , where the curve representing the end of the experiment
and the curve representing the end of life cross over one another.

2.11.2.2 Temperature/Humidity Demarcation Maps

To access the GUI for simple analytical temperature/humidity demarcation
maps, left-click on the “Temperature/humidity demarcation map” entry in
the demarcation map drop-down menu. This brings up the default GUI for
the function “array.rhplot,” shown in Figure 2.16 filled in to produce Figure 2.7.

Slot for name of object to

store numerical output Vector of experimental
Vector of times (hours)
life times
(hours) Vector of
experimental
temperatures
Vector of life (°C)
temperatures
(°C) Vector of
experimental
relative
humidities
Vector of life
relative
humidities
(1=100%)
Exponents for
Vector of log 10 (ν) relative humidity
values power laws

FIGURE 2.16
Annotated GUI for temperature/humidity demarcation map function in Splus, filled in to
produce Figure 2.7.

© 2004 by CRC Press, LLC

C4711_C02.fm Page 59 Saturday, February 28, 2004 10:29 AM

Left-clicking the apply button with the mouse produces the plot below, as
well as assigning the numerical output to the name “dum.” The output is
printed below.
The numerical output is a list; at the highest level it has two components,
corresponding to each exponent. Within each list is a sublist with compo-
nents BHT (the Teller model RH/(1-RH)), VP (the vapor pressure model),
and PL (the power law in relative humidity model). Even though there is
no numerical difference between the last two if a translation is made for
activation energy and ν, both are shown because there is a possibility that
one is interpretable when the other is not because, for example, one results
in a negative activation energy. Each of the sublists has a cc component and
a matrix component. CC shows the crossover value of log 10 ( ν) ; while the
matrix is similar to that for temperature, only the life matrix and experiment
matrix are shown side by side. The column of log 10 ( ν) values is repeated
at the beginning of both matrices, life first.

> dum
[[1]]:
[[1]]$BHT:
[[1]]$BHT$cc:
[1] 3.250845

[[1]]$BHT$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.5907987 0.6745286 2 0.6639237
[2,] 3 0.6618967 0.7381145 3 0.7360148
[3,] 4 0.7329948 0.8018350 4 0.8081058
[4,] 5 0.8040929 0.8657280 5 0.8801969
[5,] 6 0.8751910 0.9298391 6 0.9522880
[6,] 7 0.9462890 0.9942218 7 1.0243790
[7,] 8 1.0173871 1.0589357 8 1.0964701
[8,] 9 1.0884852 1.1240441 9 1.1685611
[9,] 10 1.1595832 1.1896091 10 1.2406522
[10,] 11 1.2306813 1.2556852 11 1.3127433
[11,] 12 1.3017794 1.3223126 12 1.3848343
[12,] 13 1.3728775 1.3895116 13 1.4569254
[13,] 14 1.4439755 1.4572791 14 1.5290164
[14,] 15 1.5150736 1.5255895 15 1.6011075

[[1]]$VP:
[[1]]$VP$cc:
[1] 3.78138

[[1]]$VP$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.5148744 0.5953111 2 0.5803748
[2,] 3 0.5859725 0.6589516 3 0.6524659
[3,] 4 0.6570705 0.7227423 4 0.7245569
[4,] 5 0.7281686 0.7867244 5 0.7966480
[5,] 6 0.7992667 0.8509471 6 0.8687391
[6,] 7 0.8703648 0.9154666 7 0.9408301

© 2004 by CRC Press, LLC

C4711_C02.fm Page 60 Saturday, February 28, 2004 10:29 AM

[7,] 8 0.9414628 0.9803449 8 1.0129212

[8,] 9 1.0125609 1.0456453 9 1.0850123
[9,] 10 1.0836590 1.1114273 10 1.1571033
[10,] 11 1.1547570 1.1777402 11 1.2291944
[11,] 12 1.2258551 1.2446159 12 1.3012854
[12,] 13 1.2969532 1.3120648 13 1.3733765
[13,] 14 1.3680513 1.3800736 14 1.4454676
[14,] 15 1.4391493 1.4486078 15 1.5175586

[[1]]$PL:
[[1]]$PL$cc:
[1] 9.711457

[[1]]$PL$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.5322204 0.6585367 2 0.5918326
[2,] 3 0.6033184 0.7217828 3 0.6639237
[3,] 4 0.6744165 0.7850594 4 0.7360148
[4,] 5 0.7455146 0.8483762 5 0.8081058
[5,] 6 0.8166126 0.9117461 6 0.8801969
[6,] 7 0.8877107 0.9751859 7 0.9522880
[7,] 8 0.9588088 1.0387170 8 1.0243790
[8,] 9 1.0299069 1.1023669 9 1.0964701
[9,] 10 1.1010049 1.1661696 10 1.1685611
[10,] 11 1.1721030 1.2301669 11 1.2406522
[11,] 12 1.2432011 1.2944085 12 1.3127433
[12,] 13 1.3142991 1.3589513 13 1.3848343
[13,] 14 1.3853972 1.4238571 14 1.4569254
[14,] 15 1.4564953 1.4891895 15 1.5290164

[[2]]:
[[2]]$BHT:
[[2]]$BHT$cc:
[1] NA

[[2]]$BHT$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.6443588 0.6751473 2 0.7327161
[2,] 3 0.7154569 0.7411473 3 0.8048071
[3,] 4 0.7865550 0.8076939 4 0.8768982
[4,] 5 0.8576530 0.8748106 5 0.9489892
[5,] 6 0.9287511 0.9424978 6 1.0210803
[6,] 7 0.9998492 1.0107327 7 1.0931714
[7,] 8 1.0709472 1.0794733 8 1.1652624
[8,] 9 1.1420453 1.1486638 9 1.2373535
[9,] 10 1.2131434 1.2182418 10 1.3094446
[10,] 11 1.2842415 1.2881442 11 1.3815356
[11,] 12 1.3553395 1.3583119 12 1.4536267
[12,] 13 1.4264376 1.4286923 13 1.5257177
[13,] 14 1.4975357 1.4992406 14 1.5978088
[14,] 15 1.5686337 1.5699198 15 1.6698999

[[2]]$VP:
[[2]]$VP$cc:
[1] NA

© 2004 by CRC Press, LLC

C4711_C02.fm Page 61 Saturday, February 28, 2004 10:29 AM

[[2]]$VP$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.4925103 0.5188190 2 0.5656183
[2,] 3 0.5636084 0.5852944 3 0.6377094
[3,] 4 0.6347064 0.6523382 4 0.7098004
[4,] 5 0.7058045 0.7199536 5 0.7818915
[5,] 6 0.7769026 0.7881209 6 0.8539825
[6,] 7 0.8480006 0.8568001 7 0.9260736
[7,] 8 0.9190987 0.9259368 8 0.9981647
[8,] 9 0.9901968 0.9954691 9 1.0702557
[9,] 10 1.0612948 1.0653337 10 1.1423468
[10,] 11 1.1323929 1.1354707 11 1.2144378
[11,] 12 1.2034910 1.2058267 12 1.2865289
[12,] 13 1.2745891 1.2763559 13 1.3586200
[13,] 14 1.3456871 1.3470203 14 1.4307110
[14,] 15 1.4167852 1.4177892 15 1.5028021

[[2]]$PL:
[[2]]$PL$cc:
[1] 7.424668

[[2]]$PL$mat:
matx mat.ret vec.ret matx vecx
[1,] 2 0.5272022 0.6350115 2 0.5885339
[2,] 3 0.5983002 0.6983460 3 0.6606250
[3,] 4 0.6693983 0.7617391 4 0.7327161
[4,] 5 0.7404964 0.8252093 5 0.8048071
[5,] 6 0.8115945 0.8887802 6 0.8768982
[6,] 7 0.8826925 0.9524812 7 0.9489892
[7,] 8 0.9537906 1.0163494 8 1.0210803
[8,] 9 1.0248887 1.0804294 9 1.0931714
[9,] 10 1.0959867 1.1447735 10 1.1652624
[10,] 11 1.1670848 1.2094409 11 1.2373535
[11,] 12 1.2381829 1.2744946 12 1.3094446
[12,] 13 1.3092810 1.3399970 13 1.3815356
[13,] 14 1.3803790 1.4060041 14 1.4536267
[14,] 15 1.4514771 1.4725584 15 1.5257177

2.11.2.3 Mechanical Cycling Demarcation Maps

To access the GUI for mechanical cycling demarcation maps, left-click on the
“Coffin Manson demarcation map” entry in the demarcation map drop-
down menu. This brings up the default GUI for the function “demarc.CM.”
In Figure 2.17 we show the filled in GUI that produced Figure 2.8. Left-
clicking the apply with the mouse results in Figure 2.8, and the output below
stored in the vector “dum” in Splus.

dum
$life:
matx mat.ret
[1,] 1e-005 -1.3333333 -0.18911360
[2,] 1e-004 -1.0000000 0.06088972
[3,] 1e-003 -0.6666667 0.31089688

© 2004 by CRC Press, LLC

C4711_C02.fm Page 62 Saturday, February 28, 2004 10:29 AM

Corresponding
vector of plastic
Name of object that numerical strains during
output is assigned to experiment
Name for
Vector of cycles plot (in
during life quotes)

Corresponding
vector of plastic
strains during
life Vector of M1/2
values for plot

Vector of cycles
during experiment Maximum value of –1/z for plot

FIGURE 2.17
GUI for Coffin–Manson demarcation map function in Splus for mechanical and thermal cycling,
filled in to produce Figure 2.8.

[4,] 1e-002 -0.3333333 0.56091229

[5,] 1e-001 0.0000000 0.81094548
[6,] 1e+000 0.3333333 1.06101696
[7,] 1e+001 0.6666667 1.31117079
[8,] 1e+002 1.0000000 1.56150148
[9,] 1e+003 1.3333333 1.81221051
[10,] 1e+004 1.6666667 2.06372253
[11,] 1e+005 2.0000000 2.31691025
[12,] 1e+006 2.3333333 2.57347557
[13,] 1e+007 2.6666667 2.83640022
[14,] 1e+008 3.0000000 3.10991211
[15,] 1e+009 3.3333333 3.39794514
[16,] 1e+010 3.6666667 3.70143618

$expt:
matx mat.ret
[1,] 1e-005 -1.3333333 -0.98620244
[2,] 1e-004 -1.0000000 -0.65286910
[3,] 1e-003 -0.6666667 -0.31953577
[4,] 1e-002 -0.3333333 0.01379756
[5,] 1e-001 0.0000000 0.34713090
[6,] 1e+000 0.3333333 0.68046423
[7,] 1e+001 0.6666667 1.01379756
[8,] 1e+002 1.0000000 1.34713090
[9,] 1e+003 1.3333333 1.68046423
[10,] 1e+004 1.6666667 2.01379756
[11,] 1e+005 2.0000000 2.34713090
[12,] 1e+006 2.3333333 2.68046423
[13,] 1e+007 2.6666667 3.01379756
[14,] 1e+008 3.0000000 3.34713090
[15,] 1e+009 3.3333333 3.68046423
[16,] 1e+010 3.6666667 4.01379756

>

© 2004 by CRC Press, LLC

C4711_C02.fm Page 63 Saturday, February 28, 2004 10:29 AM

The first column of both the life and the expt. matrices are the vectors of
M 1 z values; the remaining columns are the –1/z values computed from the
demarcation equations.

© 2004 by CRC Press, LLC