JOURNAL OF AIRCRAFT

Vol. 53, No. 3, May-June 2016

Deciding Optimal Number of Fatigue Crack Growth

Tests for Damage-Tolerant Design

K. S. Bhachu,∗ R. T. Haftka,† and N. H. Kim‡

University of Florida, Gainesville, Florida 32611
DOI: 10.2514/1.C033690
Fatigue crack growth coupon tests play an important role in designing safe and lightweight aircraft structures. How
many fatigue crack growth coupon tests are enough is an important question to be answered. With few coupon tests,
there will be a substantial uncertainty in design weight, with a significant chance of getting a heavier design, and a
considerable chance of overestimating fatigue crack growth design life. A simulation study is presented to determine
the optimal number of fatigue crack growth coupon tests by trading off testing cost against structural weight cost. It
also gives a designer an idea on how increasing the number of tests reduces the risk of overestimating design life. The
tradeoff is demonstrated for the damage-tolerant design of the lower spar caps of a business jet’s wing. With eight
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

coupon tests, there is a 26% chance of incurring at least 2.4% weight penalty and 18% chance of overestimating the
fatigue crack growth life by more than 10%. For 64 tests, the chances reduce to 3 and 0.5%, respectively.

Nomenclature
a = crack length, in. Superscripts
C = Paris constant
* = true parameter value
Csct = cost of single coupon tests, U.S. dollars
ˆ = estimated parameter value
Ct = cost of testing, U.S. dollars
c = effective crack length, in.
d = hole diameter, in. p I. Introduction
K = stress intensity, ksi in.
ls
Nd
=
=
length of wing spar, in.
fatigue crack growth design life, flight hours M ATERIAL testing is a key task that approximates material
properties (e.g., yield strength, crack growth rate) needed for
designing safe and low-weight aircraft structures [1]. In general,
Nt = number of load cycles in a coupon test, cycles
n = Paris exponent material properties are random and can be modeled with statistical
na = number of aircraft distributions [2]. For static strength design, the Federal Aviation
nct = number of fatigue crack growth coupon tests Administration (FAA) requires the use of A-basis or B-basis material
p = cost penalty for a pound of weight, U.S. dollars per allowable to compensate for material variability and sampling
pound uncertainty arising from limited number of coupon tests [3]. The use
t = thickness, in. of B-basis material properties could lead to heavier designs if a small
W lc = weight of lower cap, lb number (e.g., less than 30) of tests are performed. As a result, aircraft
W wing = weight of wing due to lower caps of six spars, lb companies end up performing a sufficiently large number of coupon
w = width, in. tests (more than 30) to characterize the strength property of a given
μ = mean material, which usually results in sufficient weight savings by
σ = standard deviation reducing sampling uncertainty [4,5].
In contrast, for damage-tolerant (DT) designs [6] (i.e., design
Subscripts based on fatigue crack growth life), the FAA does not mandate the use
of basis values to compensate for material variability and sampling
a = aircraft uncertainty due to the smaller number of fatigue crack growth (FCG)
d = design coupon tests. Furthermore, FCG coupon tests tend to be more time-
s = spar consuming and expensive (about $1000–2000 per test) than material
lc = lower cap strength coupon tests (about $300 per test). Consequently, aircraft
t = test companies may be performing fewer FCG coupon tests than needed.
t − ow = thickness of overweight samples It seems that some of the fatigue literature provides useful guidelines
W − lc = weight of lower caps [7–9] to determine the sufficient number of fatigue tests for under-
standing crack initiation behavior but not propagation behavior. This
is perhaps due to the concern for larger scatter in the crack initiation
life than crack growth life [10]. However, the question of how many
Received 26 August 2015; accepted for publication 16 December 2015;
FCG coupon tests are needed appears to be largely unaddressed, even
published online 28 March 2016. Copyright © 2015 by Kanwardeep S. though it might be of practical interest to airframe manufacturers.
Bhachu, Raphael T. Haftka, and Nam-Ho Kim. Published by the American Therefore, the main objective of this paper is to help designers in
Institute of Aeronautics and Astronautics, Inc., with permission. Copies of this deciding the optimal number of FCG coupon tests.
paper may be made for personal and internal use, on condition that the copier The current DT design practice for the metallic airframe is deter-
pay the per-copy fee to the Copyright Clearance Center (CCC). All requests ministic and is mostly based on predicting the mean FCG life (i.e.,
for copying and permission to reprint should be submitted to CCC at www. part sizing is based on mean FCG rate properties). At first glance, this
copyright.com; employ the ISSN 0021-8669 (print) or 1533-3868 (online) to may seem a risky design approach, but risk of failure is minimized
initiate your request. through the use of various conservative assumptions (some are dis-
*Ph.D. Student, Department of Mechanical and Aerospace Engineering;
[email protected]. cussed in Appendix A) in predicting the FCG design life of a
†
Distinguished Professor, Department of Mechanical and Aerospace component. In some cases, gross conservative safety factors are
Engineering; [email protected]. applied to the calculated FCG life. The tradeoff study presented in
‡ this paper assumes that components are designed using the mean
Associate Professor, Department of Mechanical and Aerospace
Engineering; [email protected]. FCG life as a design constraint. We present an example problem of
738
 BHACHU, HAFTKA, AND KIM 739

designing lower caps of wing spars to decide the optimal number of corresponding to each crack length measurement; C is the Paris
FCG tests. In general, with the increase in the number of FCG tests, constant; n is the Paris exponent; and ΔK is the range of stress
the mean design weight approximately remains the same, but there is intensity factor that drives crack growth.
substantial decrease in the uncertainty (due to finite sampling/ Fitting the Paris law to data (as shown in Fig. 1a) gives a single
number of tests) about the mean design weight. That is, the likelihood sample of C and n. Generally, these material constants are treated as
of getting a too-heavy design and the risk of underestimating the material properties that define the rate of crack growth for a particular
desired mean FCG life by some amount decrease with the increase in material. In practice, multiple tests are usually performed for a given
number of tests. The optimal number of tests is found by studying the stress ratio (e.g., R  0.1) and Paris fit to all the data gives a mean
tradeoff between costs of weight penalty (due to underestimation of value of C and n. These mean values are further used to predict a mean
FCG life) and FCG testing cost. We also find the risk of over- FCG life for a component, which serves as a constraint for sizing a
estimating the FCG life by 5 and 10%, which could also be used by component.
the designer to decide the number of tests in certain situations. However, the mean values of C and n are uncertain due to limited
The paper is organized as follows. Sections II and III briefly number of coupon tests (i.e., if the same number of tests is repeated,
introduce the FCG coupon testing and life prediction for designing one would get different mean values). This leads to uncertainty in the
components. Section IV introduces a procedure to simulate testing mean FCG life of a component and further leads to uncertainty in the
and life prediction. It also presents the distributions of predicted FCG design weight. Such an uncertainty could be substantial with smaller
life. These distributions are translated into distributions of design number of coupon tests and is often treated as epistemic uncertainty
thickness (or weight) that are given in Sec. V. Next, Sec. VI [14,15]. That is, there would be substantial chances of overestimating
extrapolates the weight penalties from simulation geometry to and underestimating the mean FCG design life by some amount.
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

aircraft wing spars. Finally, Sec. VII presents the cost tradeoff results The overestimation of life would lead to a lighter (thinner) design
to determine the optimal number of FCG coupon tests. that increases the risk of failure. On the other hand, underestimation
of life would lead to a heavier design that is also undesirable. In-
creasing the number of tests would considerably reduce the chances
II. Fatigue Crack Growth Coupon Testing of overestimating the mean FCG life, and the weight penalty due to
An important step in designing damage-tolerant aircraft structures underestimation would also reduce.
is the estimation of the crack growth behavior of a material through To study the effect of increasing the number of tests on uncertainty,
coupon testing. The testing is usually repeated at different stress we simulate testing by generating finite samples of C and n (each
ratios (e.g., R  0.1, 0.5) to capture the stress ratio effect for com- depicting a coupon test) from assumed true (fully sampled)
plete material characterization. For simplicity, we assume that probability distributions. These samples are further used to calculate
multiple tests are repeated for only one stress ratio. The estimated corresponding FCG life samples for a given loading and component
material properties are further used to predict the FCG life of a geometry. The mean of these FCG life samples serves as a mean life
component that is discussed in the next section. of a component. The mean FCG life is then used as a constraint for
The crack growth behavior of a material is quantified by measuring sizing a component, which would give a corresponding design
the rate of crack propagation from multiple FCG coupon tests. The weight. Such simulation is outlined in Sec. IV.
ASTM standard document [11] gives guidelines for FCG testing The simulation requires an assumption of true probability
procedures and data reduction techniques. The crack growth rate and distributions of C and n. These material constants are known to
stress intensity range data (da∕dN t versus ΔK) obtained via FCG exhibit strong negative correlation [16,17], and so these are usually
coupon testing is fitted with a crack growth model as shown in Fig. 1. modeled with a joint probability distribution. Annis [18] and
Many crack growth models (e.g., Walker equation, Nasgro equation) Akkaram et al. [19] modeled C with a log-normal distribution, or
have been proposed to model the crack growth rate data [12,13] but log(C) as normal, and n with a normal distribution using data from
are basically the extensions of the Paris law: 68 tests (2024-T3 aluminum alloy) performed by Virkler et al. [16].
The joint probability distribution function (PDF) parameters
da estimated from 68 tests is given in Table 1. For simulation, we
 CΔKn (1)
dN t assume these parameters to represent a true/fully sampled joint
distribution (usually unknown because it requires infinite number of
where da∕dN t is the crack growth rate measured in inches per cycle; tests) shown in Fig. 1b. The joint normal PDF is given in Eqs. (2)
a is the measured crack length; N t is the number of test load cycles and (3):

3.4
−4 joint prob. density
10 3.3
crack growth rate, da/dN (in/cycle)

3.2

3.1
−5
t

10 3

2.9
n

2.8
−6
10 2.7

2.6
test data
2.5
paris fit
−7
10 2.4
2 4 8 16 32 50 −19 −18.75 −18.5 −18.25 −18 −17.75 −17.5
stress intensity, ΔK (ksi√in) log(C)
a) b)
Fig. 1 Representations of a) Paris model fit to the crack growth rate data, and b) assumed true joint normal PDF of log
C and n.
 740 BHACHU, HAFTKA, AND KIM

Table 1 Assumed true (*) parameters of the bivariate crack at a hole) assumed in this paper, a surrogate model that replaces
normal distribution Eq. (4) is given in Appendix B. The random samples of C and n
Material True True standard True generated to simulate coupon testing are propagated through Eq. (4)
constant Distribution mean μ deviation σ  correlation ζ to estimate the samples of FCG life for the geometry shown in Fig. 3.
LogC Normal −18.20 0.328 −0.982 A stepwise simulation procedure is outlined in Table 3.
n Normal 2.872 0.165 −0.982 As a first step, we estimate the true (fully sampled) FCG design life
distribution via Monte–Carlo simulation (MCS). This is done by
computing Eq. (4) for 10 million samples of C and n generated from
  the assumed true joint PDF given in Table 1. The thickness of the part
1 z shown in Fig. 3 is sized such that the true mean FCG design life is
PlogC; n  p












exp − (2)
2πσ logC σ n 1 − ζ 2 21 − ζ 2  equal to 24,000 flight hours (FH). This gives the value of true thickness
(t  0.3663 in.) that is used to estimate the weight penalty. In general,
lifetime failure data are modeled with distributions having a positive
skew (i.e., a longer right tail). In [9,23,24], researchers modeled lifetime
logC − μlogC 2 2ζlogC − μlogC n − μn  n − μn 2 data with a log-normal distribution. We also model the FCG design life
z≡ −  distribution with a log-normal distribution as shown in Fig. 4. The
σ 2logC σ logC σ n σ 2n
parameters of the true FCG life’s log-normal PDF given in Eq. (5) are
(3) μln Nd  location  10.044 and σ ln Nd  scale  0.289, where μln Nd 
and σ ln Nd  are the mean and standard deviation of the natural logarithm
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

of N d :
 
III. Design of Lower Wing Spar Caps Using Fatigue 1 ln N d  − μln Nd  2
fN d  p





exp − (5)
Crack Growth Material Properties N d σ ln Nd  2π 2σ ln Nd  2
The samples of material constants generated to simulate testing are
used for calculating mean FCG design life for the wing spars/beams The mean FCG life is
shown in Fig. 2. Wing spars are primarily designed to take bending
 
loads due to aerodynamic lift, which subjects the lower spar caps σ 2ln Nd 
(e.g., shown in Fig. 2b) to axial tensile loads. Cracks often originate at μNd  exp μln Nd   (6)
2
fastener holes drilled to attach spar caps and wing skin and grow
under cyclic flight loads to a length leading to fracture.
In the viewpoint of DT design, parts (e.g., a portion of the spar cap ΔP
shown in Fig. 3) are sized such that a particular design life goal is w
achieved (e.g., 24,000 flight hours for a typical business jet spar cap
[20–22]). The FCG design life is predicted by executing an analytical
crack growth analysis for the given load conditions (e.g., using inputs
given in Table 2). For simplicity, the loading is assumed to be of w/2
constant amplitude type and number of load cycles is assumed equal
to flight hours. The load range ΔP value is selected such that the
a
thickness of the part (shown in Fig. 3) is similar to the real wing spar
cap (shown in Fig. 2b) that was designed by considering variable-
amplitude loading. For illustration, we use the integral form of the r=d/2
Paris law to calculate the design life:
c

Zaf  −n
1
Nd  ΔK da (4)
C
ai

where N d is the FCG design life, ai is the length of assumed

preexisting flaw at a hole, and af is the crack length at which failure is Fig. 3 Lower spar cap at critical fastener location subjected to axial
assumed to occur. For the part and crack geometry (through thickness tensile load.

Front Spar
Upper-cap
Web

Main Spar

Fastener

Wing Skin

Rear Spar
Lower-cap

a) b)
Fig. 2 Representations of a) wing assembly of a business jet, and b) cross-sectional view of wing assembly at main spar.
 BHACHU, HAFTKA, AND KIM 741

true distribution, μ * = 24,000 FH, t* = 0.3663" coupon tests. This reduces the chances of overestimating and the
Nd
chances of underestimating the FCG design life considerably.
0.9999 lognormal fit For example, consider the samples to the left of the true mean
0.9995 data
0.999 FCG life in Fig. 5a; for a fixed chance (e.g., 40 and 20%) of
0.995
underestimating the FCG life (shown by horizontal arrows in Fig. 5a),
0.99 the corresponding mean life increases with the increase in the number
of tests. That is, the 20th percentile and 40th percentile mean FCG life
0.95 values increase (see Fig. 5b) with increase in nct , which would
0.9
eventually lead to the decrease in the corresponding values of design
weight (discussed in the next section). On the other hand, considering
probability

0.75
samples to the right of the true mean FCG life in Fig. 5a, the
0.5 probability of overestimating the true mean FCG life of 24,000 FH
(e.g., at least by 5%, or 1200 FH, and 10%, or 1400 FH, shown by
0.25
vertical lines in Fig. 5a) can be seen to decrease rapidly with the
0.1 increase in the number of tests in Fig. 6. For example, the probability
0.05 of being greater than 25,200 FH is about 35% for nct  8 tests and
about 10% for nct  64 tests.
0.01
0.005
0.001 V. Distributions of Design Thickness
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

0.0005
0.0001 The mean FCG design life values estimated in Fig. 5a are either
10 4 10 5
larger or smaller than the desired FCG design life goal of 24,000 FH.
FCG design life, N d FH Thus, to check the effect of limited testing on design weight,
Fig. 4 Probability plot indicating log-normal distribution as a good fit to thickness of the spar cap shown in Fig. 3 is changed in a way that each
the FCG design life samples obtained via MCS. sample of μ^ Nd in Fig. 5a is equal to 24,000 FH. That is, the thickness
for the geometry shown in Fig. 3 is found such that μ^ Nd ≥ 24;000 FH.
This gives the distributions of design thickness, which are shown as
Note that, for t  0.3663 in., the true mean FCG life is exactly empirical CDFs in Fig. 7. It can be noticed that all the CDFs pass
equal to the design goal of 24,000 FH. Thus, for a very large number of through the true design thickness of 0.3663 in. at about 50%
FCG tests, one would get design weight corresponding to the true probability. The spread about true design thickness represents the
thickness of t  0.3663 in. uncertainty in the design thickness/weight values. That is, one could
obtain a design from a range of possible design weights depending on
the sample set of C and n obtained from coupon testing. However,
IV. Simulation of Crack Growth Testing and spread/uncertainty decreases with the increase in the number of tests.
Distributions of Mean Design Life In Fig. 7a, the samples of thickness to the left of the true thickness
In reality, only a handful of FCG coupon tests could be performed represent designs that are lighter due to overestimation of FCG life as
that introduce uncertainty in the mean values of C and n, which was shown in Fig. 5a. That is, the spar cap is thinner than it should be,
further lead to uncertain in the mean FCG design life value. This which would increase the risk of failure. However, it is not easy to
advances to uncertainty in the design thickness t or design weight W. estimate how much of this decrease in design weight (due to
For example, if one repeats nct number of coupon tests multiple times, overestimation of life) would offset the cost of fatigue problems due
the mean FCG design life based on repeated tests would be different to increased risk of failure. On the other hand, the thickness samples
each time, which would lead to different design weights. However, to the right of the true thickness are heavier due to underestimation of
the uncertainty would decrease with the increase in the number of FCG life. It is easier to analyze the cost benefit due to reduction in
tests. To illustrate the effect of such uncertainty, the simulation weight penalty due to these heavier designs with the increase in
procedure outlined in Table 3 is undertaken. number of tests against the testing cost. Thus, the optimal number of
The simulation gives the distributions of mean FCG design life μ^ Nd tests will be based on the mean of all the samples to the right of the
that are shown as empirical cumulative distribution function (CDFs) true thickness (0.3663 in.) shown in Fig. 7a. The mean of these
in Fig. 5a for different nct . It can be noticed that all empirical CDFs overweight thickness samples (μt−ow , horizontal line in Fig. 7a)
approximately pass through the true mean life μNd  24;000 FH at decreases with the increase in the number of tests, which is also
shown in Fig. 7b. The thickness penalty Δt also reduces, which is
about 50% probability, and uncertainty/spread about te mean shows
measured with respect to true thickness of t  0.3663 in.:
the possible values of μ^ Nd that could be obtained for a fixed true
thickness of t  0.3663 in. (estimated in the previous section), Δt  μt−ow − t (7)
depending on the sample set of C and n obtained from FCG coupon
testing. In Fig. 5a, the samples to the right of true mean life The thickness penalties corresponding to Fig. 7b (i.e., for the
overestimate the mean FCG design life, which would lead to lighter geometry shown in Fig. 3) are given in Table 4. Also, from Table 4, it
designs but would erode the margin of safety. On the other hand,
samples to the left of the true mean would lead to heavier designs,
which is also undesirable. Furthermore, uncertainty about the true Table 3 Simulation of FCG coupon testing and calculation
mean FCG life decreases rapidly with the increase in the number of of mean FCG design life
Step Description
Table 2 Dimensions and loads for the geometry shown 1 Generate nct random samples of C and n from the true joint PDF given
in Fig. 3 in Table 1.
2 Calculate samples of FCG design life N d using Eq. (4), and estimate
Dimension Value parameters ^μln N d  ; σ^ ln Nd   by fitting the log-normal distribution to the
Width w 2 in. FCG samples.
Diameter d 0.20 in. 3 Use parameters^μln N d  ; σ^ ln Nd  to estimate mean FCG design life
Initial crack length ai 0.05 in. using Eq. (6), i.e., μ^ N d .
Failure crack length af 0.90 in. 4 Repeat steps 1–3 10,000 times by generating new sample sets of C and
Load range ΔP 1 kips n to get a distribution of^μN d .
Stress ratio R 0 5 Repeat steps 1–4 for nct  f8; 16; 32; 64; 128; 256; 512; 1024g.
 742 BHACHU, HAFTKA, AND KIM

4
ECDFs of mean life, thickness t= 0.3663 in. x 10 t = 0.3663"
1 2.4

0.9
Pr.(μ Nd > 26,400 FH)

0.8 2.35
nct = 8

(FH)
cumulative probability

0.7
Pr.(μ Nd > 25,200 FH) nct = 16

Nd
0.6 nct = 32 2.3

est. mean FCG life, μ

0.5 nct = 64
Pr.( μNd ≤ μ Nd40% )= 40%
nct = 128
0.4 2.25
nct = 256
0.3 Pr.(μ Nd ≤ μ Nd20% )= 20% nct = 512

0.2 nct = 1024 2.2

th
0.1 μ *Nd (true) 20 prctl.
th
40 prctl.
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

0 2.15
1.8 2 2.2 2.4 2.6 2.8 3 8 16 32 64 128 256 512 1024
est. mean FCG life, μ (FH) x 10
4
number of coupon tests, n ct
Nd
a) b)
Fig. 5 Representations of a) empirical CDFs of mean FCG life for t  0.3663 in., and b) 20th and 40th percentile mean FCG life as a function of number
of coupon tests.

 
can be noticed that the probability of being heavier by more than 1 1 9
2.4% (shown by the vertical line in Fig. 7a) reduces with the increase μW−lc  wlc ls μt−ow  wlc ls μt−ow ρal  wlc ls μt−ow ρal
2 16 16
in nct . For example, for nct  8, there is 26% chance of being heavier
by more than 2.4%, i.e., Prt > 0.375 in.), which reduces to 3% (8)
for nct  64.
where ρal  0.105 lb∕in.3 As nct → ∞, μt−ow → t  0.3663 in.,
and μW−lc → W lc  24.92 lb. This is about 70% of the weight of the
VI. Extrapolation of Weight to Wing Spar and Fleet real main spar’s lower cap shown in Fig. 2. Next, a typical business
Typically, for a business jet’s wing, approximately 50–70% of the jet’s wings have six spars, and so their combined weight is
wing spar’s lower-cap sizing is dominated by the FCG life constraint. approximated using the following equation:
The thickness results listed in Table 4 are extrapolated to calculate the
weight penalty for the entire lower cap of a spar. Thus, we consider μW−wing  2μW−ms  μW−fs  μW−rs   21  0.50  0.75μW−lc
that the wing spar is ls  24 ft long (a typical length of a midsized 9
business jet spar [20]), and for the first half of spar’s length, the lower  μW−lc (9)
2
cap’s thickness is assumed to be t and width as wlc  4 in. Further,
for the second half of the spar’s length, the thickness and width are where μW−ms is the mean weight of the main spar, μW−fs is the mean
assumed to be 1∕4t and wlc ∕2. The mean weight for entire lower cap weight of the front spar (about 50% of the main spar), and μW−rs is the
of a single spar corresponds to the mean thickness of overweight mean weight of the rear spar (about 75% of the main spar). Therefore,
samples (μt−ow ): the weight penalty for the wings due to six spar caps (ΔW wing ) is
calculated as follows:
true mean life, μ * = 24,000 FH
10 0
Nd
ΔW wing  μW−wing − W wing (10)

where W wing is the true weight of the wing due to six lower
10 −2 caps corresponding to the true thickness of t  0.3663 in.
(i.e.W wing  112.2 lb). The corresponding weight penalties are
prob. of over−estimating true life by β %

10 −4
listed in Table 5. It can be noticed that the weight penalty is about
3.53 lb (about 3% of W wing ) for nct  8 and reduces to about 0.3 lb
(0.3% of W wing ) for 1024 tests.
10 −6

10 −8
VII. Cost of Weight Penalty and Testing Cost Tradeoff
The weight penalties calculated in the previous section are based
on the mean thickness of the overweight samples (from Fig. 7a) can
10−10 now be used for calculating the cost penalty to the operators due to
additional weight attributable to finite coupon testing. The worth of
β = 5%, Pr(μ Nd > 25,200 FH) structural weight proposed by Curran et al. [25] is 300∕kg (136∕lb).
10−12 Similarly, Kim et al. [26] referred to the U.S. National Materials
β = 10%, Pr( μ Nd > 26,400 FH) Advisory Board report [27] that estimated the worth of weight saved
as 200∕lb for a civil transport aircraft. Acar et al. [5] varied the cost of
10−14
8 16 32 64 128 256 512 1024 weight penalty p between 200∕lb and 1000∕lb. Bhachu et al. [21]
number of coupon tests, n ct proposed a measure of worth based on useful load and found that p
Fig. 6 Probability of overestimating FCG design life by 5 and 10% as a varies between $800 and $1600 for business jets. In this paper, we
function of number of tests. vary p between 200∕lb and 1200∕lb to see its effect on the optimum
 BHACHU, HAFTKA, AND KIM 743

ECDFs of thickness (t), est mean FCG design

mean FCG life (μ Nd = 24,000 FH)
life = 24,000 FH
1 0.38
μ t−ow
0.9
μ t−ow
cumulative probability
0.8 0.3775

n ct = 8
0.7
Pr.(t < 0.375") 0.375

thickness, t (in.)
n ct = 16
0.6
n ct = 32

0.5 n ct = 64 0.3725

0.4 n ct = 128

n ct = 256
0.37
0.3
n ct = 512

0.2 n ct = 1024
0.3675
0.1 t true = 0.3663"

0 0.365
0.34 0.35 0.36 0.37 0.38 0.39 0.4 8 16 32 64 128 256 512 1024
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

thickness, t (in.) ct number of coupon tests, n

a) b)
Fig. 7 Representations of a) CDFs of thickness, and b) mean of overweight thickness samples as a function of number of tests.

number of FCG coupon tests. The cost associated with weight The tradeoff results are shown in Fig. 8. It can be noticed from
penalty is calculated using the following formula: Fig. 8a that, for weight penalty of p  200∕lb , the optimal number
of FCG coupon tests is around 64 (i.e., the point where both costs are
CΔW−fleet  na ΔW wing p (11) balanced), and for p  1200∕lb, it is about 210. Thus, one could
justify 64–210 FCG tests for designing damage-tolerant components
where na is the number of aircraft (400), i.e., the weight penalty is (e.g., wing spar caps). On the other hand, if the designer is too
transformed into the cost penalty for the entire fleet of aircraft to be concerned about eroding the margins of safety on the FCG design
produced by a manufacturer. On the other hand, the cost of coupon life, then Fig. 6 could be used to qualitatively decide the number of
testing increases linearly with the number of FCG coupon tests: tests. For example, if the designer is okay with the 1% risk of
overestimating the FCG life by 5%, then the optimal number of tests
Ct  nct Csct (12) would range from 128 to 256 (refer to Fig. 6). However, if the
designer is fine with 1% risk of overestimating the desired mean FCG
where Csct is the cost of single coupon tests that ranges between life by 10%, then the optimal number of tests would range from 32 to
$1000 and $2000. These cost numbers are based on personal 64. In any case, the current amount of FCG coupon testing that
communications with engineers at Cessna Aircraft Company. We typically ranges from 8 to 32 seems to be somewhat inadequate.
assume the mean cost of $1500 per test for our analysis. Notice that The results presented in this paper are based on the assumption of
the cost of FCG coupon testing is much higher than $300 for material the Paris law, but in reality, models that take stress ratio into account
yield strength tests mentioned in [5]. are employed (e.g., Walker equation). In those cases, the testing is
repeated for multiple stress ratios, and so the optimal number of tests
could be distributed equally among those testing conditions. For
Table 4 Mean thickness of overweight samples, example, if testing is repeated at four different stress ratios, then the
thickness penalties, and probability of overdesigning optimal number of tests (e.g., 64) may be distributed among four
stress ratios (i.e., 16 tests per stress ratio). However, such an
nct μt−ow , in: Δt, in: Prt > 0.375 in., %
assumption should be verified by performing a simulation similar to
8 0.3778 0.0115 26 the one presented in this paper with the improved model.
16 0.3743 0.0080 18
32 0.3719 0.0056 9.5
64 0.3702 0.0039 3
128 0.3690 0.0027 0.6 VIII. Conclusions
256 0.3683 0.0020 0.03 From the preceding analysis, it was demonstrated that performing
512 0.3676 0.0013 — — more FCG coupon tests reduces the risk of both overestimating the
1024 0.3672 0.0009 — —
→∞ 0.3663 — — — —
mean FCG design life and overdesigning the structure, although it is
not easy to decide the optimal number of tests by calculating the cost
associated with the reduction in the chances of fatigue problems (due
to overestimation of life). But one could make an informed decision
Table 5 Weight penalty for wing spars based on mean by looking at an acceptable probability (e.g., 1%) of overestimating
thickness of overweight samples the life by some percentage (e.g., 5 or 10%). Such an approach may be
useful for deciding the optimal number of tests for designing damage-
nct μW−wing , lb ΔW wing , lb tolerant components (e.g., stringers) that does not drive much weight
8 115.7 3.53 into the airframe.
16 114.6 2.45 On the other hand, for structures like wing spars, a decision on the
32 113.9 1.71 optimal number of tests could be made by studying the tradeoff
64 113.4 1.20 between costs of weight penalty and testing cost. It was shown that
128 113.0 0.82 about 64–256 coupon tests could be justified in comparison to the
256 112.8 0.61
current practice of about 8–32 tests, depending upon the associated
512 112.6 0.41
1024 112.4 0.29
weight penalty that was assumed to range from $200/lb to $1200/lb.
→∞ 112.2 — — Therefore, we conclude that the current amount of FCG coupon
testing that ranges from 8 to 32 could be increased to maximize the
 744 BHACHU, HAFTKA, AND KIM

7
10 10 7
weight−cost weight−cost
test−cost test−cost

6
10 10 6
Cost (log10 −scale), $

Cost (log −scale), $

5
10 10 5

10
4
10 10 4

3
10 10 3
8 16 32 64 128 256 512 1024 8 16 32 64 128 256 512 1024
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

number of coupon tests (log 2−scale), n ct number of coupon tests (log 2−scale), nct
a) b)
Fig. 8 Tradeoff between weight penalty cost and testing cost for a) p  200∕lb , and b) p  1200∕lb.

chances of weight savings and reducing the chances of fatigue where Fwc is the finite width correction, and G∞
cf is the infinite plate
problems due to overestimation of life. solution. The integrand in Eq. (B1) requires numerical integration
that is computationally expensive in the simulation. Therefore, it is
approximated by the surrogate given in Eq. (B3) that is valid for the
Appendix A: Conservative Assumptions of Crack Growth geometry given in Table 2:
Analysis
  −n
The deterministic crack growth analysis makes many conservative p



assumptions (implicit safety factors) and explicit safety factors for N d  tn C−1 ΔP π ∕w
conservatively predicting the FCG design life. Some implicit safety  
factors/conservative assumptions are as follows. × 1.64 − 1.22n  1.14n − 0.3n  0.047n
2 3 4
B3
1) Conservative assumption about the size of preexisting rogue
flaw (e.g., 0.05 in.). The probability of such a flaw to exist in a pristine
structure is about 10−6 [28].
2) Mean crack growth rate curves are multiplied with a safety References
factor to derive the upper bounds. [1] Acar, E., Haftka, R. T., and Kim, N.-H., “Effects of Structural Tests on
3) The material testing is usually performed under a high-humidity Aircraft Safety,” AIAA Journal, Vol. 48, No. 10, 2010, pp. 2235–2248.
and high-temperature environment that is known to accelerate the doi:10.2514/1.J050202
crack growth rate. Thus, material constants obtained by fitting crack [2] Ocampo, J., Millwater, H., Singh, G., Smith, H., Abali, F., Nuss, M.,
growth models to such data are conservative. Reyer, M., and Shiao, M., “Development of a Probabilistic Linear
4) Crack growth models often only model the linear Paris region Damage Methodology for Small Aircraft,” Journal of Aircraft, Vol. 48,
(i.e., where crack growth rates are between 10−6 and 2 × 10−5 inches No. 6, 2011, pp. 2090–2106.
per cycle, or between ΔK  6 − 16 ksip
in.). That is, ignoring near- doi:10.2514/1.C031463
[3] Composites Materials Handbook MIL-HDBK-17-1F, Vol. 1, Depart-
threshold crack growth data tends to underpredict the crack growth ment of Defense Handbook, Composite Materials Handbook, 1997,
life. pp. 14–85, Chap. 8.
5) Sometimes a factor of safety (e.g., 1.5 to 2) is used to knock [4] Park, C., Kim, N. H., and Haftka, R. T., “How Coupon and Element
down the calculated FCG life. Tests Reduce Conservativeness in Element Failure Prediction,”
Therefore, in the presence of such conservative assumptions, one Reliability Engineering and System Safety, Vol. 123, March 2014,
hopes to cover for the uncertainty arising from flight load spectrum pp. 123–136.
and material randomness. doi:10.1016/j.ress.2013.10.012
[5] Acar, E., Haftka, R. T., and Kim, N.-H., “Reliability-Based Structural
Design of Aircraft Together with Future Tests,” 51 st AIAA/ASME/
Appendix B: Fatigue Crack Growth Life Based on ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Geometry, Load Conditions, and Stress Intensity Solution Conference, AIAA Paper 2010-2595, April 2010.
[6] Goranson, U. G., “Damage Tolerance: Facts and Fiction,” Keynote
The FCG life equation given in Eq. (4) can be expanded as follows:
Presentation at International Conference on Damage Tolerance of
Aircraft Structures, Delft Technical Univ., Delft, The Netherlands,
  Zaf  
p


−n p


 2007.
Nd  tn C−1 ΔP π ∕w Gcf a da (B1) [7] Machin, A. S., How Many Specimens? An Aid in the Design of Fatigue
Test Programs, Dept. of Defence Paper AD-A14281, Melbourne,
ai
Australia, 1984, pp. 1–6.
    2 [8] Gope, P. C., “Determination of Minimum Number of Specimens in S-N
r r Testing,” Journal of Engineering Materials and Technology, Vol. 124,
Gcf  Fwc G∞  0.7071  0.7548  0.3415
cf
ra ra No. 4, 2002, pp. 421–427.
 3  4  [9] Gope, P. C., “Determination of Sample Size for Estimation of Fatigue
r r Life Using Weibull or Lognormal Distribution,” International Journal
 0.642  0.9196
ra ra of Fatigue, Vol. 21, No. 8, 1999, pp. 745–752.
s
























































    doi:10.1016/S0142-1123(99)00048-1
πr πr  a∕2 [10] Schijve, J., “Fatigue Predictions and Scatter,” Fatigue and Fracture of
× sec sec (B2) Engineering Materials and Structures, Vol. 17, No. 4, 1994, pp. 381–396.
w w−a doi:10.1111/j.1460-2695.1994.tb00239.x
 BHACHU, HAFTKA, AND KIM 745

[11] “Standard Test Method for Measurement of Fatigue Crack Growth Rates,” Vol. 52, No. 5, 2015, pp. 1412–1421.
ASTM International Rept. ASTM E647-13ae1, West Conshohocken, PA, doi:10.2514/1.C032945
2013. [21] Bhachu, K. S., Haftka, R. T., Kim, N. H., and Hurst, C., “Performance of
[12] Toor, P. M., “A Review of Some Damage Tolerance Design Approaches Different Alloys by Assessing Epistemic and Aleatory Uncertainty in
for Aircraft Structures,” Engineering Fracture Mechanics, Vol. 5, No. 4, the Crack Growth Rates,” 16th AIAA Non-Deterministic Approaches
1973, pp. 837–880. Conference, AIAA Paper 2014-1009, 2014.
doi:10.1016/0013-7944(73)90054-4 [22] Bhachu, K. S., Haftka, R. T., and Kim, N. H., “Comparison of Methods
[13] Miller, M. S., and Gallagher, J. P., “An Analysis of Several Fatigue Crack for Calculating B-Basis Fatigue Crack Growth Life Based on Limited
Growth Rate (FCGR) Descriptions,” Fatigue Crack Growth Measurement Coupon Tests,” AIAA Journal, Vol. 48, No. 6, 2011, pp. 2090–2106.
and Testing, ASTM International Rept. STP-738, 1981, pp. 14–85. doi:10.2514/1.C031463
doi:10.1520/STP738-EB [23] Yang, J. N., and Manning, S. D., “Aircraft Fleet Maintenance Based on
[14] Thacker, B., Riha, D., Millwater, H., and Enright, M., “Errors and Structural Reliability Analysis,” Journal of Aircraft, Vol. 31, No. 2, 1994,
Uncertainties in Probabilistic Engineering Analysis,” 19th AIAA Applied pp. 419–425.
Aerodynamics Conference, AIAA Paper 2011-1239, 2011. doi:10.2514/3.46502
[15] Bae, H.-R., Grandhi, R., and Canfield, R. A., “Uncertainty Quantification [24] Yang, J. N., and Manning, S. D., “Stochastic Crack Growth Analysis
of Structural Response Using Evidence Theory,” AIAA Journal, Vol. 41, Methodologies for Metallic Structures,” Engineering Fracture
No. 10, 2003, pp. 2062–2068. Mechanics, Vol. 37, No. 5, 1990, pp. 1105–1124.
doi:10.2514/2.1898 doi:10.1016/0013-7944(90)90032-C
[16] Virkler, D. A., Hillberry, B. M., and Goel, P. K., “The Statistical Nature [25] Curran, R., Rothwell, A., and Castagne, S., “Numerical Method for Cost-
of Fatigue Crack Propagation,” U.S. Air Force Flight Dynamics Lab., Weight Optimization of Stringer–Skin Panels,” Journal of Aircraft,
Rept. AFFDL-TR-78-43, 1978. Vol. 43, No. 1, 2006, pp. 264–274.
[17] Ostergaard, D. F., and Hillberry, B. M., “Characterization of the
Downloaded by UNIVERSITY OF FLORIDA on June 20, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.C033690

doi:10.2514/1.13225
Variability in Fatigue Crack Propagation Data,” Probabilistic Fracture [26] Kim, H. A., Kennedy, D., and Gurdal, Z., “Special Issue on Optimi-
Mechanics and Fatigue Methods, ASTM International Rept. STP-798, zation of Aerospace Structures,” Structural and Multidisciplinary
West Conshohocken, PA, 1983, pp. 97–115. Optimization, Vol. 36, No. 1, 2008, pp. 1–2.
[18] Annis, C., “Probabilistic Life Prediction Isn’t as Easy as It Looks,” doi:10.1007/s00158-008-0256-1
ASTM International Rept. STP-1450, Vol. 1, No. 2, West Conshohocken, [27] “Materials Research to Meet 21st Century Defense Needs,” AD-
PA, 2004. A142810, National Research Council, Melbourne, Australia, 2003,
[19] Akkaram, S., Loghin, A., and Khan, G., “A Framework for Probabilistic http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&
Assessment of Fracture Mechanics Life,” AFRL-RX-WP-TP-2011- identifier=ADA142810.
4390, Niskayuna, NY, 2011. [28] Millwater, H. R., and Wieland, D. H., “Probabilistic Sensitivity Based
[20] Bhachu, K. S., Haftka, R. T., Waycaster, G., Kim, N.-H., and Hurst, C., Ranking of Damage Tolerance Analysis Elements,” Journal of Aircraft,
“Probabilistic Manufacturing Tolerance Optimization of Damage Vol. 47, No. 1, 2010, pp. 161–171.
Tolerant Aircraft Structures Using Measured Data,” Journal of Aircraft, doi:10.2514/1.44498