J Mater Sci (2012) 47:2770–2781

DOI 10.1007/s10853-011-6106-3

A new statistical framework for the determination of safe creep

life using the theta projection technique
M. Evans

Received: 30 August 2011 / Accepted: 2 November 2011 / Published online: 18 November 2011
Ó Springer Science+Business Media, LLC 2011

Abstract In this article a new estimation framework is rupture represents an obvious life limiting design consid-
put forward for the well-known theta projection technique eration as fracture of blades used in aero engines could
which enables, for the first time, levels of confidence to be prove catastrophic. For this reason, studying the ability to
associated with the creep property predictions made using predict the time to rupture strain has been the major criteria
this technique. The predictions made from the resulting used to assess creep extrapolation techniques. However,
model are in the form of distributions, which is a sub- substantial problems can also be encountered due to
stantial advance on existing life assessment methodologies excessive creep distortion. There are numerous examples
used in high-temperature applications such as the disks and of such deformation limits within the power generation and
blades used in aero engines. This additional information aero engine industries. For example, the blades of a steam
should prove invaluable for questions related to the issues turbine cannot be allowed to extend until they foul the
of possible life extension. When applied to data on surrounding casting. Similar requirements exist for the
Ti.6.2.4.6, accurate interpolations and extrapolations could blades used within a gas turbine engine.
be made of the actual distribution of the minimum creep Any realistic creep property projection technique needs
rate, which in combination with the Monkman–Grant to contain the following features. First, there needs to be an
relation, also enabled accurate predictions to be made for experimental stage where constant stress creep curves are
the time to failure. For example, when extrapolating to a measured over a range of accelerated stresses and tem-
stress of 480 MPa, the time to failure was predicted to peratures. Second, the resulting experimental creep curves
follow a non-normal distribution with a median time of need to be modelled in such a way that they can be pro-
3450 h and a 95% confidence interval of 3250–3800 h. The jected to other stresses and temperatures—either within
single experimental data point available at this stress was (i.e., interpolation) or outside (i.e., extrapolation to the
consistent with such an extrapolation. service conditions) the original range of test conditions.
Finally, it must be possible to ‘‘read off’’, in some way, the
required creep properties (such as the minimum creep rate
Introduction or the time to x% strain) from the projected creep curve.
In contrast to traditional parametric procedures, (such as
When designing materials for high-temperature service the the Larson–Miller technique [1]), the H projection tech-
design criteria for long-term operation must guarantee that nique allows the whole creep curve to be extrapolated to
creep deformation should not cause excessive distortion design (low) stresses from accelerated stresses. The tradi-
over the planned service life and that creep failure should tional theta projection technique, as developed by Evans
not occur within such a required operating life. Such creep and Wilshire [2], achieves this in the following way.
Extensive research on this technique has shown that

M. Evans (&) e ¼ H1 ð1  eH2 t Þ þ H3 ðeH4 t  1Þ ð1aÞ

Materials Research Centre, College of Engineering,
Swansea University, Swansea SA2 8PP, UK gives a good representation (at least for large strains) of
e-mail: [email protected] experimental creep curves for many steel alloys [3, 4],

123
J Mater Sci (2012) 47:2770–2781 2771

nickel-based super alloys [5], and other aerospace materials 0.1%) before a planned replacement date. This article aims
[6]. In Eq. 1a, e is the recorded creep strain at time t and Hh to overcome this limitation by updating the theta prediction
are unknown parameters. H1 and H3 scale the primary and technique in the light of the latest developments in the
tertiary components of creep, respectively, whilst H2 and statistics of parameter estimation in engineering applica-
H4 are the corresponding rate constants governing these tions. In particular, the two step procedure is replaced by a
two components of creep. Their values depend upon the unified single step approach where predictions are naturally
test conditions under which the experimental creep curves in the form of a distribution of values for a creep property
were obtained. Again Evans and Wilshire proposed that rather than a single value. To illustrate this new statistical
this test dependency could be represented by framework for the theta projection technique, data on
ln(Hh Þ ¼ bh1 þ bh2 s þ bh3 T þ bh4 sT; ð1bÞ Titanium 6.2.4.6 is used.
To achieve this objective the article is structured as
where s is stress, T temperature and bh1 to bh4 further follows. The next section briefly reviews the data set to be
unknown parameters. At very small strains, the fit of Eq. 1a used in illustrating this new statistical framework. This is
to the experimental data is sometimes quite poor. This followed by a section outlining the new statistical model,
misspecification has lead to considerable errors in the together with some of the salient stochastic features of creep
calculation of times to small strains and to initial and that is must have built into it. Details of how to estimate this
minimum creep rates in various titanium [7] and aluminium model are left to an attached Appendix. Section 4 applies
alloys [8]. In such circumstances, the following the unified estimation methodology to the data on Ti.
generalisation of Eq. 1a was put forward by Evans [9] 6.2.4.6 and then conclusions are drawn in a final section.
e ¼ H1 ð1  eH2 t Þ þ H3 ðeHt  1Þ þ H5 ð1  eH6 t Þ ð1cÞ
where the first two terms on the right hand side have the Experimental procedures and the test matrix
same physical meaning as in Eq. 1a. The first term on the
right hand side of this equation describes normal primary The material used to illustrate the new unified estimation
creep and the second term normal tertiary creep. The third methodology is the Titanium alloy Ti-6.2.4.6 prepared as an
term describes early primary behaviour. ingot forged in the b-phase at approximately 1238 K (the b
Evans [10] has described the detailed numerical and transus is approximately 1223 K). The chemical composi-
statistical techniques required to estimate all the unknown tion of this material (in wt%) was determined as 5.79Al,
parameters contained within the above equations. He 1.99Sn, 3.94Zr, 6.03Mo, 0.06Fe, 0.02C, 0.1O, 0.0035Si and
developed a sophisticated two step procedure for applying Balance *82Ti. The initial heat treatment schedule used
the theta projection technique. First, values for the Hh was 1173 K for 2 h, followed by air quenching. It was then
parameters are estimated by applying a non linear least reheated to 868 K, held for 8 h, and finally air quenched. A
squares algorithm to Eq. 1a. Then, and using these esti- second re-age for 2 h at 913 K, in air, was employed, which
mates, a weighted least squares procedure is used to esti- is a simulated post weld heat treatment.
mate the values for bh1 to bh4 in Eq. 1b. Weighted, rather Twenty one conventional creep specimens, from mate-
than straightforward linear least squares, is required rial supplied by TIMET U.K. Ltd., of 3.8 mm diameter,
because the Hh values are estimated from the experimental 25.4 mm gauge length, and 3/8 inch BSF thread were
data, i.e. they are measured with error. Consequently, the machined from the heat treated material and tested in ten-
weights used should be equal to the inverse of the variances sion over a range of stresses at 773 K using seven high-
for Hh. Estimating these variances is further complicated precision uniaxial constant-stress machines fitted with three
by the fact that the scatter of experimental date points zone furnaces. All machines had been calibrated to British
around fits given by Eq. 1a often appear to follow a random Standards BSEN10002 (parts 1–5). Details of such testing
walk. Failure to take this into account results in a sub- machines can be found in most texts on creep [11]. Tem-
stantial under estimate of these variances. perature was maintained along the gauge length and with
The use of such a two step procedure makes it very respect to time to better than ±0.5 K and the extensometers
difficult to obtain creep property predictions with confi- were capable of establishing creep strains to better than
dence limits attached to them. That is, whilst this two step 10-5 and the creep strain-time curves contained up to 400
procedure is capable of producing a point prediction for points. Digital data collection was used in all cases.
properties such as the time to failure or the minimum creep The test matrix used is summarised in Table 1. Thirteen
rate, it cannot easily produce estimates of the distributions of the twenty one specimens were placed on test at 773 K
for these properties. Yet, such distributions are critical for and at a stress of 580 MPa, with the remaining eight
determining how long a material should stay in service specimens being tested at 773 K and over the stress range
such that the chances of failure are minimal (e.g., less than 900–480 MPa—excluding 580 MPa. The creep properties

123
2772 J Mater Sci (2012) 47:2770–2781

Table 1 Isothermal (773 K) testing matrix showing the number of 0.2

specimens tested on various test rigs and at different stresses Tests carried out at 580 MPa
Test carried out at 535 MPa
Stress Test rig number 0.15

Strain, ε (%/100)
(MPa)
1 2 3 4 5 6 7
0.1
900 1
800 1
700 1 0.05

620 1
600 1
0
580 2 3 3 3 2 0.0E+00 1.0E+06 2.0E+06 3.0E+06 4.0E+06 5.0E+06 6.0E+06

560 1 Time, t (s)

535 1 Fig. 1 Uniaxial creep curves obtained at 773 K:580 MPa and at
480 1 773 K:535 MPa

The creep of Ti. 6.2.4.6 is therefore a highly stochastic

contained in this Table have been reported (in part) else- process. Indeed, this stochastic component of creep is just
where and more details on this data can be found in Evans as important, if not more important, than the deterministic
and Ward [12]. element which describes how the shape of the creep curve
varies with the test conditions. This is again visualised in
Fig. 1 where the creep curve obtains at 535 MPa (a change
A new statistical framework for the theta projection of nearly 50 MPa) is within the scatter band of the creep
technique curves obtained at 580 MPa! Yet this stochastic element of
creep receives only scant attention in the academic litera-
Salient features ture. It is therefore very important to have a suitable sta-
tistical framework for dealing with such a large variability
Between variation in repeat measurements.

By this it is meant that specimens put on test under the same Within variation
uniaxial conditions for stress and temperature will produce
different shaped creep curves and therefore different mini- By this it is meant that the strain time readings made on any
mum creep rates and times to failure. The source of this one specimen are measured with error so that the experi-
variation depends on the creep database being used. In the mental creep curve will be scattered about the curve defined
test program described in Sect. 2 above it reflects differing by any equation chosen to model creep strain. To look at
test rigs as all specimens were cut from the same batch of this second salient feature in more detail, it is necessary to
material. In the larger ECCC and NIMS databases [13, 14], generalise Eq. 1a by adding to it a random variable, v, to
the between variation comes from different test rigs, dif- pick up this within variation (or indeed to Eq. 1c, although
ferent batches of material, different heat treatments, and this article will limit its scope to a four theta analysis)
sometimes different laboratories. The first salient feature    
eij ¼ H1 1  eH2 tij þ H3 eH4 tij  1 þ vij ð2Þ
therefore that needs to be captured by any suitable statistical
framework is that the parameters of the equation chosen to where eij is the actual strain measurement made on speci-
model the experimental creep curve obtained at a given test men i at time tij (so the times need not be the same for all
condition must come from some type of joint distribution. test specimens) and vij is the residual variation for speci-
Figure 1 shows the 13 experimental creep curves obtained men i at time tj. The total number of inspections on spec-
at 580 MPa and 773 K for Ti-6.2.4.6 and as can be seen there imen i is denoted by ni. The specimens tested at 580 MPa
is substantial variation between the results obtained on each were labelled i = 1–13 for this paper. The variable v, being
of the test rigs. Even at the same test condition the rupture a random variable, will follow some type of distribution to
strain and the time at which each specimen failed were very be discussed further below.
different—for example the rupture strain varied over the The second salient feature that needs to be captured in
range 7–15%. These observed variations in the experimental any suitable statistical framework is that the random vari-
creep curves shown in Fig. 1 must by definition be random in able vij follows a random walk. To visualise this, consider a
nature. It follows from this that the theta values in Eq. 1a single specimen and a single test condition, e.g., specimen
must therefore also be random variables. i = 1. (In Table 1 this corresponds to the first specimen

123
J Mater Sci (2012) 47:2770–2781 2773

0.025
0.02 (a)
0.015
0.01
0.005
v1j

0
-0.005
-0.01
-0.015
-0.02
-0.025
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
141
146
151
156
161
166
171
176
181
186
191
196
201
206
211
216
221
226
231
236
241
246
251
256
261
266
271
276
281
286
j
0.025

v1j
(b) 0.02
0.015
0.01
0.005
0
-0.025 -0.02 -0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02 0.025
-0.005
v1j-1
v1j= 1.002v1j-1 -0.01
R² = 98.2%
-0.015
-0.02
-0.025

Fig. 2 Variation of v1j with a the observation number and b it immediate previous value

placed on test on test rig 1 at 580 MPa and 773 K). Some eij ¼ H1 ð1  eH2 tij Þ þ H3 ðeH4 tij  1Þ þ qvij1 þ uij ð4aÞ
n1 = 284 strain–time recording were made on this speci-
men during the test. If H1 to H4 in Eq. 2 are then chosen to where q = 1 and uij is a new random variable free from
P 1 ¼284 2 this autocorrelation. Rearranging Eq. 2 for vij and
minimise nj¼1 v1j the following result is obtained
    substituting into Eq. 4a gives
e1j ¼ 0:0157 1  e3:004E05t1j þ 0:1138 e2:324E07t1j  1
eij  qeij1 ¼ H1 ðPij  qPij1 Þ þ H3 ðTij  qTij1 Þ þ uij
þ v1j ð3Þ
ð4bÞ
Figure 2 plots the values for v1j as calculated from
where Pij and Tij stand for the primary and tertiary com-
Eq. 3. It is clear from this figure, that v1j and v1j-1 for all j
ponents of creep, respectively, i.e., Pij ¼ ð1  eH2 tij Þ; Pij1
are highly dependent. When v1j is plotted out against j, as
in Fig. 2a, a low frequency wave pattern is observed. Then ¼ ð1  eH2 tij1 Þ; Tij ¼ ðeH4 tij  1Þ; Tij1 ¼ ðeH4 tij1  1Þ.
when v1j is plotted against its immediate previously Figure 3a shows the u1j values for the same test specimen
recorded value, the correlation coefficient is q = 1.002 as in Fig. 2 (again using non-linear least squares estimates
as shown in Fig. 2b. When this correlation coefficient of the theta values shown in Eq. 3 and when q = 1). As
equals one, the strain series is then said to follow a random can be seen the signature tune for first order autocorrelated
walk—i.e., the first difference of v1j are then completely errors is no longer present (the data points now appear
random in nature. This random walk phenomenon is not random with no cyclical or alternating variation).
specific to this test specimen but can be observed in all test
specimens (no matter what the test condition is) when Heteroscedasticity
Eq. 1a is fitted to the experimental data associated with
that test condition (it is also true for other materials as By this it is meant that scatter present in the measurement
well). This phenomenon was first identified by Evans [10] errors associated with the experimental creep curves for each
in his article discussing some of the statistical aspects of specimen are not the same. The third salient feature that
the theta projection technique. Given that this correlation needs to be captured in any suitable statistical framework is
coefficient is approximately 1 in value, a solution to this that the variability in the random variable u in Eq. 4b varies
problem is found by working in first differences. Writing between the experimental creep curves. This is visualised in
Eq. 3 to allow for this type of first order autocorrelation Fig. 3b where the uij values associated with all i = 1, 13
gives, specimens tested at 580 MPa (again using non-linear least

123
2774 J Mater Sci (2012) 47:2770–2781

Fig. 3 a The variation of u1j 0.006

with the observation number
(a)
0.004
and b the spread in values for uij
with specimen number, with all 0.002
specimens tested at 580 MPa

u1j
0
and 773 K
-0.002

-0.004

-0.006

1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
134
141
148
155
162
169
176
183
190
197
204
211
218
225
232
239
246
253
260
267
274
281
j
0.015
(b)
0.01

0.005
uij

-0.005

-0.01

-0.015
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Specimen i

squares to estimate the theta values for each specimen and 1

Normal distribution (mean = 0,
using q = 1) are displayed. What is clear from this figure is 0.9 standard deviation = 0.0012)
Empirical distribution
that the scatter present in the random variable u is not con- 0.8

stant between the test specimens. 0.7

Probability

0.6

Distributions 0.5

0.4

The final salient feature that needs to be captured in any 0.3

suitable statistical framework is that the random variable u 0.2
in Eq. 4b will follow a distribution and there is no a prior 0.1
reason for assuming this to be the normal distribution. This 0
-0.006 -0.004 -0.002 0 0.002 0.004 0.006
is visualised in Fig. 4 where the empirical cumulative
u1j
density function for the u1j values for the same test spec-
imen as that shown in Fig. 2 is given. This empirical Fig. 4 The empirical cumulative distribution for u1j, together with
density function is found by sorting all the u1j from lowest the cumulative normal distribution for a zero mean variable with a
to highest to form the sorted series u1j(k), where k is the standard deviation of 0.0012
position of u1j in the sorted list. Each value for u1j(k) is then
given an empirical cumulative probability equal to (k-
0.5)/n1, where n1 is the number of observations on u1j. Also eij  qeij1 ¼ H1i ðPij  qPij1 Þ þ H3i ðTij  qTij1 Þ þ gi uij

shown in this figure is the normal cumulative density with uij ¼ uij gi
function drawn for a standard deviation of 0.0012 (which is ð5aÞ
the sample standard deviation for u1j). It is immediately
apparent that there is a difference between the actual with q = 1 and with
empirical distribution and the normal distribution so that it vij ¼ qvij1 þ uij ; Pij ¼ ð1  eH2i tij Þ;
is important to build into the model the possibility of non-
Pij1 ¼ ð1  eH2i tij1 Þ; Tij ¼ ðeH4i tij  1Þ;
normally distributed errors so that any divergence from a
normal distribution can at least be tested for. Tij1 ¼ ðeH4i tij1  1Þ

The first salient feature described in sub Sect. 3.1 is

The statistical model captured in Eq. 5a by allowing H1–H4 to be different for
each of the i specimens put on test. The third salient feature
The starting point is Eq. 4b, but with the random error is captured by allowing g to be different for each of the i
termed scaled by a further parameter gi, and with all the specimen tested. This is so because, and as shown below,
theta parameters varying with each specimen tested the value of this scale parameter is determined by the

123
J Mater Sci (2012) 47:2770–2781 2775

scatter present in the strain measurements around each Lawless [20] for a good review of this distribution). The
fitted creep curve. This randomness in the theta and scaling probability density function for u*ij is then given by
values can be further modelled by specifying a joint    
kk0:5 pffiffiffi uij þ D
distribution for them—which in this article is taken to be f ðuij Þ ¼ exp kfuij þ Dg  kexp pffiffiffi
CðkÞ k
the joint normal distribution. This appears quite a
reasonable assumption to make because if maximum ð5dÞ
pffiffiffi
likelihood techniques are used to estimate these where D = k ½wðkÞ  ln(kÞ with w(k) being the di
parameters, the resulting estimates are asymptotically gamma function [21], ensuring that u*ij has a mean of zero
normal anyway. Some restrictions on this joint normal at all values for k. k determines how u*ij is actually dis-
distributions seem appropriate, however. As the scatter in tributed. As k ? ?, u*ij tends to a normal distribution, but
H1i–H4i picks up the between test rig variability in this when k = 1, u*ij has an extreme value distribution. The
article, whilst the scatter in the gi values pick up the within exponential and gamma distributions are also special cases,
machine variability in the strain measurements, it seems but for all other k values, u*ij follows the modified log
appropriate to assume these two sources of variation are gamma distribution. The value for k is chosen so that the
independent of each other. The mean vector and covariance modelled distribution is a best fit to the empirical distribu-
matrix of this restricted joint normal distribution are then tion. To see how Eqs. 5a, 5d also deals with the third salient
given by l and R, respectively, where feature described above, take some special cases. When
2 3
lH1 k = ?, gi is the standard deviation of u*ij, whilst when
6 lH2 7 k = 1, gi is 1.283 divided by this standard deviation. Hence
6 7
l¼6 7
6 lH3 7 and the scatter present in the strain measurements making up the
4 lH4 5 experimental creep curves, as measured by the standard
lg deviation, can be different for each test specimen.
2 2 3 ð5bÞ
rH1 rH12 rH13 rH14 0 For a programme of testing that just puts multiple
6 rH12 rH2 rH23 rH24 0 7
2
specimens on test at a single stress–temperature combina-
6 7
R¼6 2
6 rH13 rH23 rH3 rH34 07 7 tion, the model defined by Eqs. 5a–5d is complete. How-
4 rH14 rH24 rH34 r2H4 0 5 ever, when multiple specimens are put on test at a variety
0 0 0 0 r2g of different conditions, some modification of the above
model is required to take into account the change in creep
where lH1 is the average of all the H1i values (i.e., averaged
curve shape that occurs when stress and temperature
over all specimens tested), lH2 is the average of all the H2i
change. Generally speaking, the primary component
values and so on up to lg which is the mean value for all the
diminishes with decreasing stresses and temperatures. This
gi values. r2H1 is the variance of H1, r2H2 is the variance of
additional variation can be taken into account using
H2 and so on up to r2g which is the variance of g. rH12 is the
acceleration factors ðAF h Þ as follows
covariance between the H1 and H2 values and rH13 is the
covariance between H1i and H3i and so on. A more compact eij  qeij1 ¼ AF 1 H1i ðPij  qPij1 Þ
way of stating this is through the expression 
þ AF 3 H3i ðTij  qTij1 Þ þ gi uij with uij ¼ uij gi
li ¼ l þ zi ð5cÞ vij ¼ qvij1 þ uij ; Pij ¼ ð1  eAF 2 H2i tij Þ;
where the new vector variable zi follows a joint normal Pij1 ¼ ð1  eAF 2 H2i tij1 Þ; Tij ¼ ðeAF 4 H4i tij  1Þ;
distribution with a mean vector of zero and the same Tij1 ¼ ðeAF 4 H4i tij1  1Þ ð6aÞ
covariance matrix as in Eq. 5b. At this stage it is useful to
realise that this type of model has a very wide sphere of where AF h (h = 1, 4) are the acceleration factor defined
applicability and are therefore known by different names in such that AF = 1 when the stress corresponds to the stress
the academic literature. In the social sciences they are you wish to extrapolate to, but AF [ 1 at stresses above
known as multilevel models (see Goldstein [15] for a these. Typically, this extrapolated stress would correspond
review), in the statistical and econometrics literature as to the conditions under which the material normally oper-
random effect models (see for example Laird and Ware ates at (in an aero engine for example). The advantage of
[16] and Lindstrom and Bates [17] and Madalla et al. [18]) this representation is that the estimates of the Hh parame-
and in the lifetime data analysis literature as shared frailty ters obtained using accelerated test data, should correspond
models (see Haggard [19] for an excellent review). to those estimates that would be made if non accelerated
The second salient feature is accounted for by fixing test data was available and used.
q = 1 in Eq. 5a. To account for the fourth salient feature, In this article, only one varying stress is present—the
u*ij is taken to follow a generalised gamma distribution (see load on the specimen—as temperature is held constant.

123
2776 J Mater Sci (2012) 47:2770–2781

Typical applications of the theta projection technique have with each test specimen. This is done by simulating values
used the following functional form for AF h for Hh using the estimated values for l and R and
AF h ¼ exp½ch s  where s ¼ ð sU  sÞ ð6bÞ weighting them by the log likelihood of observing all the
strain time reading for specimen i given these simulated
and where sU, is the in use stress (or stress level you wish to vales for Hh. By repeating this process many times a
extrapolate too), whilst s, is an accelerated stress. ch are weighted average for the Hh values can be obtained, where
characteristic of the particular degradation process leading to simulated values for Hh that are inappropriate for specimen
eventual failure. The format of Eqs. 6a, 6b can easily be i (but may be more appropriate for another specimen)
adapted to incorporate programmes with changing receive a low weight. This weighted average for the Hh
temperatures as well. The ch parameters are treated as parameters then enables a predicted creep curve to be
common across all test specimens, as these are a constant drawn for specimen i. This weighted average technique
material property associated with the material under will be called a best predictor.
investigation. (For example, if s was a thermal stress, then c2
and c4 would be the activation energies associated with the rate
parameters H2 and H4—the activation energies for primary and Results
tertiary creep processes). Consequently, Eq. 5b is amended to
23 At a single test condition
lH1
6 lH2 7 To achieve numerical stability the strain v time data on
6 7
6 lH3 7 each specimen tested were rescaled to lie in the range of
6 7
6 lH4 7 zero through to unity, by diving the strain measurements by
6 7
l¼6 7
6 lg 7 and the rupture strain and the times by the failure time of that
6 lc1 7
6 7 specimen. All parameter estimates shown below are
6 lc2 7
6 7 therefore in these scaled units. SML techniques were used
4 lc3 5
to estimate the parameters of Eq. 5 using the scaled data
lc4
2 2 3 associated with those test specimens placed on test at
rH1 r2H12 r2H13 r2H14 0 0 0 0 0 580 MPa only (specimens i = 1 to 13). This was done for a
6 r2 r2H2 r2H23 r2H24 0 0 0 0 07
6 H12 7 variety of different k values in Eq. 5d and the maximised
6r 2
r2H23 r2H3 r2H34 0 0 0 0 07
6 H13 7 log likelihood was recorded in each instance. When k = 1,
6 r2 2
r2H34 r2H4 07
6 H14 rH24 0 0 0 0 7 the maximised log likelihood value was 19101, when
6
R¼6 0 0 0 0 r2g 0 0 0 077 k = 5 it was 19798, when k = 10 it was 19986, when
6 0 0 0 0 0 0 0 0 07
6 7 k = 20 it was 20104 and when k = 1000 it was 20280.
6 0 0 0 0 0 0 0 0 07
6 7 These results suggest that u* ij in Eq. 5a follows a standard
4 0 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 normal distribution. Thus the normal distribution for uij
was selected in all the subsequent analysis. The SML
ð6cÞ
estimates for the parameters in Eq. 5b, 5c, when k = ?,
The Appendix to this article explains how all the were found to be
parameters in this modified theta projection technique can 2 3
0:072
be estimated using simulated maximum likelihood 6 7
6 47:106 7
techniques (SML). Knowing the values for l and R 6 7
allows random drawings to be made for the values of H1– l¼6 7
6 0:502 7 and
6 7
H4 and from this simulated creep curves can be drawn for 4 1:044 5
any test condition (given also the estimates made for lch). 0:0013
By obtaining many such simulations a distribution of creep 2 3
3:3E - 05 0:0119 0:0002 0:0002 0
curves for any test condition can be derived, and from these 6 7
6 0:0119 33:798 0:0288 0:0417 0 7
simulated curves a distribution of other creep properties 6 7
R¼6 6 0:0002 0:0288 0:0050 0:0066 0 7
can be easily calculated, such as the minimum creep rate 7
6 7
distribution and times to a particular strain distribution. To 4 0:0002 0:0417 0:0066 0:0088 0 5
get a failure time distribution, a failure criteria must 0 0 0 0 0:0004
additionally be specified such as a critical rupture strain or
perhaps the Monkman–Grant relation [22]. The magnitude of r2g = 0.0004, suggests that the within
Whilst the Hh parameters are random variables, it is also variation is approximately constant between specimens, so
possible to obtain an estimate of the Hh values associated that there appears to be little or no heteroscedasticity for this

123
J Mater Sci (2012) 47:2770–2781 2777

0.16 1
Empirical distribution at 580 MPa
0.14 0.9 Modelled distribution
0.8
0.12
Strain, ε (%/100)

0.7

Probability
0.1
0.6
0.08 0.5

0.06 0.4
0.3
0.04
0.2
0.02 Experimental creep curves at 580 MPa
Best predictor 0.1
0
0 1000000 2000000 3000000 4000000 5000000 6000000 0
-5 -4 -3 -2 -1 0 1 2 3
Time, t (s)
Standardised values for the natural log
Fig. 5 Uniaxial creep curves obtained at 773 K and 580 MPa, of the minimum creep rate (s -1)
together with the best predictor for each specimen
Fig. 6 The standardised empirical and predicted (log) minimum
Titanium alloy. Further, the magnitude of r2H2 = 33.798, creep rate distribution for specimens tested at 773 K and 580 MPa
suggests that rate parameter governing primary creep is by
far the most variable of all the theta parameter. More 1

informative, however, is the correlation matrix associated 0.9

with R. That matrix is equal to 0.8

2 3 0.7
1
6 0:36 7
Probability

6 1 7
0.6
6
R ¼ 6 0:37 0:07 1 7 0.5
7
4 0:37 0:08 0:99 1 5 0.4
0 0 0 0 1 0.3

0.2
It is clear from this correlation matrix that the theta Empirical distribution at 580 MPa
0.1
values of Eq. 5a are far from independent of each other— Modelled distribution

an implicit assumption made when estimating the theta 0

-3 -2 -1 0 1 2 3 4 5
values in the traditional way using the method proposed by Standardised values for the natural log of times to failure (s)
Evans. There is moderate correlation between H1 and the
other three theta parameters, virtually no correlation Fig. 7 The standardised empirical and predicted (log) failure time
distribution for specimens tested at 773 K and 580 MPa
between H2 and H3 or between H2 and H4, but almost
perfect (and negative) correlation between H3 and H4. minimum creep rates (-17.73) and the dividing the out-
Figure 5 again replicates the creep curves measured for come by the standard deviation (0.32) in the 13 measured
the 13 specimens put on test at 580 MPa, together with the log minimum creep rates). Superimposed on this figure is
best predictors of these 13 specimens obtained in the way the predicted distribution obtained by calculating the
described at the end of the previous sub section. It is clear minimum creep rate from 2000 simulated creep curves,
from this figure that the model predicts well the shape of where each simulation was obtained using the above values
each specimens experimental curve, except perhaps at the for l and R. The predicted distribution appears to be a good
lower end of the range. But even here, the tertiary com- description of the actual distribution, and both are clearly
ponent of creep is predicted quite well. long tailed and non normal.
In Fig. 6 the dots trace out the actual or empirical Figure 7 is derived in the same way as Fig. 6, but using
standardised distribution for the log of the minimum creep the 13 log failure times at 580 MPa. To get the predicted
rate. This empirical distribution is found by sorting all the failure time distribution using the above estimates, the
experimentally measured log minimum creep rates (of predicted log minimum creep rates in Fig. 6 were con-
which there are 13 at the 580 MPa test condition) from verted into predicted log failure times using ln(tzf ) = -0.97
lowest to highest to form the sorted series e_ m ðkÞ, where k is ln(_ezm ), where tzf are the experimentally derived failure times
the position of e_ m in the sorted list. Each value of e_ m ðkÞ is standardised in the same way that the minimum creep rates
then given an empirical cumulative probability equal to were standardised, and e_ zm are the experimentally derived
(k-0.5)/13. These probabilities are then plotted against the log minimum creep rates, again standardised as described
standardised value for experimentally measured minimum above. (0.97 comes from the linear regression of ln(tzf ) on
creep rate (obtained by subtracting the mean of the 13 log ln(_ezm ). The predicted distribution appears to be a good

123
2778 J Mater Sci (2012) 47:2770–2781

description of the actual or empirical distribution and most AF 1 ¼ exp½0:0032s ; AF 2 ¼ exp½0:0041s ;

interestingly the model predicts the actual long tail
AF 3 ¼ exp½0:0022s ; AF 4
apparent at the highest failure time recorded at 580 MPa.
¼ exp½0:0017s ; if s  620 MPa
Multiple test conditions AF 1 ¼ exp½0:0049s ; AF 2 ¼ exp½0:0179s 
AF 3 ¼ exp½0:0005s  AF 4 ¼ exp½0:0004s 
Bourgeois et al. [23, 24] have studied in some detail the
otherwise
deformation and damage micro mechanisms of Ti. 6.2.4.6
in tension and uniaxial creep at 773 K. They found that The results at 480 MPa were excluded during this
creep behaviour could be divided into three domains gov- estimation process as the above parameter estimates are
erned by the applied stress, and two of these regimes cor- next used to extrapolate the various creep properties to this
responded to the test conditions studied in this article. At lower stress test condition.
medium stress levels the a grains that had been strained Based on these estimated values for ch, together with
were divided into low, high, and sub-boundary dislocation those for l and R shown above, Fig. 9 shows the models
arrangements. The authors proposed that in this stress predictions for the shape of the experimental creep curves
region there was competition between the formation and obtained at 700 and 535 MPa. It can be seen from this
annihilation of sub boundaries that lead to an extended figure that at the extremes of the data used to estimate the
period of time spent around the minimum creep rate. At the acceleration factors, the model predicts the shape of the
higher stress level there were hardly any non-strained a experimental creep curve very well. Whilst accurate
grains together with a small percentage of sub-boundary interpolations of this nature are useful inputs into finite
and low density dislocation formations. The authors were
0.12
then of the opinion that the high-stress levels caused 535 MPa 700 MPa Interpolations
0.11
instability in the sub-boundaries that led toward a 0.1
homogenous arrangement of high density dislocations. 0.09
Strain, ε (%/100)

This formation in turn lead to very short primary and ter- 0.08
tiary periods with no appreciable time spent around the 0.07

minimum creep rate. This general picture is confirmed in 0.06

Fig. 8, which contains the experimental creep curves 0.05

0.04
obtained at all the other stresses shown in Table 1 (other
0.03
than at 580 MPa).
0.02
The switch from one type of creep curve shape to the 0.01
other appears to be around the 620 MPa. This phenomenon 0
0.0E+0 1.0E+06 2.0E+06 3.0E+06 4.0E+06 5.0E+06
manifested itself in the shape of different values for the ch
Time, t (s)
values of Eq. 6b above and below this stress level. Using
all the creep data in Fig. 8, except the results obtained at Fig. 9 Uniaxial creep curves obtained at 773 K and at stresses of
480 MPa, resulted in the following estimates for the 700 MPa and 535 MPa, together with the models interpolations at
these stresses
parameters in Eq. 6b

0.2

0.15
Strain, ε (%/100)

0.1

0.05

0
0.0E+00 2.0E+06 4.0E+06 6.0E+06 8.0E+06 1.0E+07 1.2E+07 1.4E+07
Time, t (s)
Fig. 10 Uniaxial creep curve obtained at 773 K and 480 MPa,
Fig. 8 Uniaxial creep curves obtained at 773 K and at stresses from together with the models extrapolations—with confidence limits—to
900 to 480 MPa (excluding test carried out at 580 MPa) this stress

123
J Mater Sci (2012) 47:2770–2781 2779

element models of deformation, for estimating the life However, the most likely time to failure, defined by the
expectancy of materials operating at high temperatures, median prediction is 3450 h—which in combination with the
extrapolative accuracy is often more important. above limits suggest the failure time distribution is non
In Fig. 10, the experimental creep curve obtained at symmetric. Confirmation of such predictions would also form
480 MPa is shown, together with the models predictions of an interesting topic for future research.
the shape of this creep curve. As this 480 MPa data was not
used in obtaining the ch values shown above these pre-
dictions are indeed extrapolations. The prediction shown in Conclusions
this figure is in fact the median prediction obtained by
simulating the model under these conditions, but also A new estimation framework has been put forward for the
shown are the estimates made for the 0.05 and 99.5 per- well-known theta projection technique which enables, for
centiles. It is interesting to note how these limits expand as the first time, levels of confidence to be associated with the
creep progresses. It is clear from this figure that whilst the creep property predictions made using this technique. The
early primary creep behaviour is not predicted very well, predictions made from the resulting model are in the form
the experimental data is within the confidence limits of distributions which is a substantial advance on existing
around the minimum rate and for all tertiary creep. It would life assessment methodologies used in high-temperature
therefore be beneficial, as part of a future research project, applications such as the disks and blades used in aero
to see whether using the 6-H methodology summarised in engines. These capabilities could also be applied to the
Eq. 1c would better predict this primary stage of creep. power generating industry if applied to suitable materials
Whilst this is not so important for lifing materials it is and thereby contribute to the future security of supply.
important for situations where times to small stains are an When this new statistical approach was applied to
important design factor. Ti.6.2.4.6 a number of observations could be drawn. First,
The model should therefore be able to predict minimum allowing the parameters describing the shape of the creep
creep rates and time to failure at this condition quite well. This curve to be random variables that follow a joint normal
is confirmed in Fig. 11, where for each property the experi- distribution, enabled an accurate representation of creep
mental data point is well within the limits of the predicted scatter at unchanging test conditions to be made. Second, it
cumulative probability distributions. Figure 11a suggests also enabled accurate interpolations to be made of the
that if more tests were carried out on this material at 480 MPa, actual distribution for minimum creep rates to be made,
there is a 95% chance that the resulting minimum creep rates which in combination with the Monkman—Grant relation,
would all be within the range 1.1E-12 s-1 to 1.3E-12 s-1. enabled accurate interpolations of the failure time distri-
Figure 11b suggests that if more tests were carried out on this bution to be made. Third, whilst extrapolation to stresses
material at 480 MPa, there is a 95% chance that the resulting lower than those used in estimation were unable to accu-
failure times would all be within the range 3250–3800 h. rately represent early primary creep, tertiary creep was
Fig. 11 Experimentally derived
values for a the time to failure
and b the minimum creep rate at
480 MPa, together with the
predicted distributions obtained
by the model

123
2780 J Mater Sci (2012) 47:2770–2781

well-predicted and consequently so too were the times to The log likelihood for all i units is then
failure. More specifically, the extrapolations to 480 MPa X m
suggested the times to failure will follow a non normal lnðLÞ ¼ lnðLi jH1i  H4i ; c1  c4 ; gi Þ
distribution with a median time of is 3450 h and a 95% i¼1 8 9
X m X ni < ðk  0:5Þ lnðkÞ  ln CðkÞ  lnðgi Þ =
confidence limit of 3250–3800 h.
¼ pffiffiffi
 pffiffi
þ k uij þ D  keðDþuij Þ= k ;

Suggestions for future research could include a study to :
i¼1 j¼1
see how much better the model performs when using a six
parameter equation rather than a four parameter equation to ð10Þ
describe creep curve shape. More replicate creep curve where there are m test specimens. In principle this log
experimentation would also further improve the parameter likelihood could be maximised with respect to H1i–H4i, c1
estimates and testing to lower stresses would further to c4, and gi. However, there are two problems with this
quantify extrapolative capabilities. Finally, the statistics approach. First, the variance–covariance matrix R and the
associated with extending the model to cover more than mean matrix l are absent from this log likelihood and so
one source of between variation could be researched. cannot be estimated in this way. This problem can be
However, as there are currently few data sets in existence overcome by specifying the random nature of H1i to H4i
where complete creep curves are available for different and gi
batches of the same material, further experimentation
would be required to verify such extended models. H1i ¼ lH1 þ z1i ; H2i ¼ lH2 þ z2i ; H3i ¼ lH3 þ z3i ;
H4i ¼ lH4 þ z4i ; gi ¼ lg þ z5i
ð11Þ
Appendix
where lH1 is the average of all the H1i values through to
The modified theta projection method takes the form given lH4 which is the average of all the H4i values and lg is the
by Eq. 6a. average of all the gi values . Further, z1i to z5i are assumed
to follow a joint normal distribution with a mean vector of
eij  qeij1 ¼ AF 1 H1i ðPij  qPij1 Þ zero and variance–covariance matrix given by R, with the

þ AF 3 H3i ðTij  qTij1 Þ þ gi uij with uij ¼ uij gi restriction that the Hh and g are independent of each other.
vij ¼ qvij1 þ uij ; Pij ¼ ð1  eAF 2 H2i tij Þ; Further, z1i to z5i can calculated from standard normal
variates in the following way
Pij1 ¼ ð1  eH2i tij1 Þ; Tij ¼ ðeAF 4 H4i tij  1Þ;
z1i ¼ dH1 w1i ; z2i ¼ dH12 w1i þ dH2 w2i ;
Tij1 ¼ ðeH4i tij1  1Þ
z3i ¼ dH13 w1i þ dH23 w2i þ dH3 w3i
ð7Þ ð12Þ
z4i ¼ dH14 w1i þ dH24 w2i þ dH34 w3i þ dH4 w4i ;
with q fixed at unity and with the AF h defined by Eq. 6b. z5i ¼ dg w5i ði ¼ 1 to mÞ
The probability density function for the change in
consecutive strain measurements made on specimen i is where w1i to w5i are standard normal variates, i.e., variables
given by that follow independent normal distributions with means of
hpffiffiffin o zero and variances of 1. Further, these new delta values are
  kk0:5
f eij  qeij1 ¼ exp k uij þ D related to the variance–covariance matrix R as follows
gi CðkÞ 2 3
   ð8Þ dH1 0 0 0 0 0 0 0 0
uij þ D
kexp pffiffiffi 6 dH12 dH2 0 0 0 0 0 0 07
k 6 7
6 dH13 dH23 dH3 0 0 0 0 0 07
6 7
6d 7
where u*ij is defined by Eq. (7). This type of distribution 6 H14 dH24 dH34 dH4 0 0 0 0 0 7
K¼6 0 6 0 0 0 dg 0 0 0 0 7
was first put forward by Bartlett and Kendall [25] and re 7
6 0 0 0 0 0 0 0 0 07
parameterised by Prentice [26] to improve stability in 6 7
6 0 0 0 0 0 0 0 0 07
estimation. Given q = 1, the log likelihood for the ni 6 7
  4 0 0 0 0 0 0 0 0 05
values of f eij  qeij1 made on specimen i, conditional
0 0 0 0 0 0 0 0 0
on H1i–H4i and c1 to c4 is then given by
with KKT ¼ R
Xni n
lnðLi jHli  H4i ; c1  c4 ; gi Þ ¼ ðk  0:5Þ lnðkÞ ð13Þ
j¼1 ð9Þ
pffiffiffi
  pffiffi o The superscript T in Eq. 13 stands for transpose so that K is
 ln CðkÞ  lnðgi Þ þ k uij þ D  ke ðDþuij Þ= k
a Cholesky decomposition of R.

123
J Mater Sci (2012) 47:2770–2781 2781

The basic steps involved in maximising this resulting References

log likelihood have been summarised by Greene [27] as
follows: 1. Larson FR, Miller J (1952) Trans ASME 174(5)
2. Evans RW, Wilshire B (1999) Introduction to creep, Appendix A.
1. Obtain at random 5 draws from the standard normal The Institute of Materials, Bournemouth
distribution using the well known Box-Muller [28] 3. Evans RW, Wilshire B (1984) Creep of metals and alloys. The
Institute of Metals, London
method. These values then quantify w1i to w5i in 4. Evans RW, Willis MR, Wilshire B, Holdsworth S, Senior B,
Eq. 12, for i = 1. Fleming A, Spindler M, Williams JA (1993) In: Proceedings of
2. Repeat steps one above m times to obtain values for the 5th international conference on creep and fracture of engi-
w12 to w52 through to w1m to w5m. neering materials and structures, University College Swansea,
Swansea, p 633
3. Choose, at random, starting values for all the param- 5. Brown SGR, Evans RW, Wilshire B (1986) Mater Sci Eng
eters in K of Eq. 13. 84(1–2):147
4. Insert the values obtained in steps 2 to and 3 into 6. Evans RW, Hull RJ (1996) J Mater Proc Technol 56(1–4):492
Eq. 12 to obtain values for z1i to z5i for i = 1 to m. 7. Evans M (2001) J Mater Sci 36:2875. doi:10.1023/A:1017946
218860
5. Choose, at random, starting values lH1 to lH4, lg and 8. Evans RW (2001) Mater Sci Technol 17(5):487
all the c parameters in Eq. 6b. Insert these, together 9. Evans RW (2000) Mater Sci Technol 16(1):6
with the z1i to z5i values obtained in step 4 above into 10. Evans RW (1989) Mater Sci Technol 5(7):699
Eq. 11 to obtain values for H1i to H4i and gi (for i = 1, 11. Penny RK, Marriott DL (1995) Design for creep. Chapman &
Hall, London
m). 12. Evans M, Ward AR (2000) Mater Sci Technol 16(10):1149
6. Insert the values obtained for H1i to H4i, gi and c1 to c4 13. Yokoi S (ed) National Institute for Materials Science, various
from step 5 above into Eq. 7 to calculate values for u*ij creep data sheets, Tokyo
(using q = 1). Then insert all these u*ij values into 14. ECCC Recommendations. In: Bullough CK, Merckling G (eds)
Data exchange and collation, vol 4, Holdsworth
Eq. 10 to obtain a numeric value for the log likelihood 15. Goldstein H (2003) Multilevel statistical models, 3rd edition,
given a value for k. appendix to chapter 2. Wiley, London
7. Then run a standard numerical optimisation procedure 16. Laird NM, Ware JH (1982) Biometrics 38:963
to maximise ln(L) with respect to lH1 to lH4, lg, c1 to 17. Lindstrom MJ, Bates DM (1988) J Am Stat Assoc 83:1014
18. Maddala GS, Li H, Trost RP, Joutz F (1997) J Bus Econ Stat
c4 and the parameters in K. (K can then be converted to 15:90
R using Eq. 13). 19. Hougaard PP (2000) Analysis of multivariate survival data.
8. Finally, steps 1–7 are then repeated R times yielding R Springer-Verlag, New York
values for lH1 to lH4, lg, c1 to c4 and the parameters in 20. Lawless JF (1982) Statistical models and methods for lifetime
data, Chapter 1. Wiley, New York
K. The average of these R values can be taken as the 21. Abramowitz M, Stegun IA (eds) Handbook of mathematical
simulated maximum likelihood estimates for these functions. Dover, New York
parameters, and the standard deviation in these R 22. Monkman FC, Grant NJ (1963) In: Mullendore AW, Grant NJ
values as an estimate of their standard error. The (eds) Deformation and fracture at elevated temperature. MIT
Press, Boston
maximised log likelihood is then given by the average 23. Bourgeois M, Feaugas X, De Mestral F, Chauveau T, Cla M
of the R maximised log likelihood values, lnðLÞ (1995) In: Proceedings of the 8th world conference on titanium,
International Convention Centre, Birmingham
This estimation procedure can be repeated using dif- 24. Bourgeois MM, Feaugas X, De Mestral F, Chauveau T, Cla M
ferent values for k, and the value for k that maximises lnðLÞ (1996) Rev Metall Cshiers Inf Tech 93(12):1521
25. Bartlett MS, Kendall DG (1946) J R Stat Soc
is chosen as the correct value for k. This is quite a
26. Prentice SRL (1975) Biometrika 61:539
straightforward procedure in practice because obtaining 27. Greene WH (2008) Econometric analysis, 6th edn. New Jersey,
random draws from a standard normal distribution is easily Prentice Hall
implemented within popular and commercially available 28. Box GEP, Muller ME (1958) Annal Math Stat 29:610
29. Microsoft Excel: Microsoft Office (2007)
software packages such as Microsoft Excel [29]. In this
package the Normsinv (Rand()) function is used to obtain a
value at random from this distribution.

123