CHAPTER 14

SENSITIVITY ANALYSIS AND UNCERTAINTY

PROPAGATION OF COMPUTATIONAL MODELS

B. F. BLACKWELL
K. J. DOWDING
Sandia National Laboratories
Albuquerque, New Mexico, USA

14.1 INTRODUCTION 443

14.2 COMPUTATION OF SENSITIVITY COEFFICIENTS 444
14.2.1 Differentialion of Analytical Solutions 444
14.2.2 Finite-difference Methods 448
14.2.3 Complex-step Method 450
14,2.4 Software-differentiation Method 451
14.2.5 Sensitivity Equation Method 452
14.2.6 Sensitivity Equation Method for Multidimensional Problems 458
14.2.7 Differentiation of Discrete Equations and Adjoint Methods for Steady
Conduction Problems 458
14.3 FIRST-ORDER PROPAGATION OF UNCERTAINTY IN COMPUTATIONAL
MODELS 460
14.4 SAMPLING METHODS FOR SENSITIVITY AND UNCERTAINTY ANALYSIS 462
14.5 SUMMARY 466

14.1 INTRODUCTION

With each advance in microelectronics, computing speed continues to risc while hardware costs
continue to drop. It is now possible to perform many morc heat transfer simulations than
were conceivable a decade ago. Instead of performing single-point calculations using nominal
parameter values (such as density, specific heat, viscosity, thermal conductivity, emittance),
it is now pOssible to consider the sensitivity of the response to parameter variations. Good
heat transfer analysts have always performed informal parameter sensitivity analyses to answer
thermal design-related questions. With the availability of tooay's computing resources, it is
now possible to take computational heat transfer to the next plateau and routinely compute
sensitivity information. It is likely that more robust designs will evolve from enhanced utilization
of computing resources.

443
444 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

We feel that a sensitivity analysis is as important as the nominal parameter calculation,

Nor only will the sensitivity analysis identify those parameters thai arc important, it will also
identify those parameters that arc unimportant. For the unimportant parameters, engineering judg-
ment may be adequate since these parameter values do not materially impact the computational
results. Resources can be focused on obtaining better estimates of the important parameters.
The sensitivity analysis can be used as a road map of how to best utilize limited resources.
Sensitivity information is also used as an integral part of any gradient-based optimization or
parameter-estimation problem (SCI! Chapter 17, The Inverse Problem).
Once sensitivity information has been computed, it is a relatively small (computational) step
to estimate the parametric uncertainty in the computational results, given estimates of uncertainty
in the various model parameters. The experimental community has been estimating uncertainty
in results from data reduction equations for many years. This procedure involves sensitivity
information in some form. The process of estimating uncertainty in computational results due to
pammetric uncertainty is analogous to estimating experimental uncertainty. Some journals have
adopted a policy of requi ring an uncertainty statement abOut presented experimental results; the
authors advocate the development. of a similar policy for purely computational results .
This chapter first addresses techniques for computing sensitivity information, followed by
methods for propagating uncertainty through the computational modeL Numerous example cal-
culations are presented. Additional details on sensitivity analysis and uncertainty propagation
can be found in Gacllci [IJ and Saltelli et al. (2J.

14.2 COMPUTATION OF SENSITIVITY COEFFICIENTS

A measure of the sensitivity of some scalar response function rex, t, p) to changes in a particular
parameter Pi is termed the sensitivity coefficient and is defined as

.. .
sen~ahvlty CDC
ffi Clent:::::
. _a'-'(",x._I:..:
. p-,-) (14.1)
api
where x is the position vector, t is time, and Pi is one element of p (the vector of all problem
parameters). The response function could be temperature, velocity, heat flllX, shear stress, etc. For
a sing le material problem involving thermal conductivity. volumetric specific heat, viscosity. and
emittance, the parameter vector would be p = {k pCp J.,L e(. Many materials will be present
for industrial heat conduction or conjugate heat transfer problems. In this case, the thermal
properties of all the materials present will be part of the parameter vector; consequently, the
parameter vector can contain tens 10 hundreds of elements. The methOds for computing sensitivity
coefficients include

• Differentiation of analytical sol utions

• Finite difference (in parameter space)
• Complex step
• Software differentiation
• Sensitivity equation method
• Adjoint method
The remainder of this section disl:usses each of these techniques.

14_2_1 Differentiation of Analytical Solutions

Differentiation of analytical soluUons is probably the simplest method of computing sensitivity
coefficients. provided that an analytical solution is available for the problem. To demonstrate
COMPUTATION OF SENSITIVfTY COEFFICIENTS 445

differentiation of analytical solutions, and to introduce some sensitivity coefficient related con-
cepts, consider transient. constant property, heat conduction in a one-dimensional planar slab
with a constant flux boundary condition. The mathematical statement of the problem is

(14.2)

- k -UT I
ax .< =0
~ q aT
- k -
ax
l x= L
~ O 1'(., , 0) ~ 1; ( 14.3)

The analytical solulion to this problem has been presented in Arpaci l3J and is

¢ (::. ~) = T - 'Ii = rtf + ~ _ :. + ~ (::)2 _ ~ ~ ~ exp (_1I2rr2 Off) cos (nrr:')

L ' L2 qL j k L2 3 L 2L rr 2 L.- n2 L2 L
,,:1
(14 .4)
where a is the thermal diffusivily (k j p c: p ). Results from this problem have also been dis-
cussed in Beck and Arnold [4J and Beck et al. (5]. The parameter vector for this problem is
p = {k p Cp q L 'Ii IT. Norc that since p Cp appears as a product , it can be defined as a single
parclmeter and is tenned the volumetric heat capacity C: this may not be the case for fluid flow
problems . II is convenient to write the temperature profile in the fonn

(14.5)

We now calculate the heat-flux sensitivity coefficient. Differentiating Eq. ( 14.5) with respect to
q yields
-~-¢
ar
L
(14.6)
uq k

Si nce aT j aq is independent of q , the temperature field is termed linear-in -the-paramerer q. The

result presented in Eq. (14.6) demonstrates that sensitivity coefficients may have uncommon
units. e.g. [temperature/parumeter unitsl. To eliminate the uncommon units, it is common to scale
the sensitivity coefficients by the nominal parameter value . The sca led sensitivity coefficient and
the corresponding dimensionless sensitivity coefficient become

aT qL T,
T ~ q- ~ -¢ Xq~ -- ~¢ ( 14.7)
q aq k qL / k
Note that the units of Tq arc temperature: this approach will allow the magnitude of the
various scaled sen sitivity coefficients to be compared directly . For simple textbook problems, it
may be convenient to use dimensionless sensitivity coefficients. However. problems involving
complt!x gcomclric.:s, lemperature-dependent properties, and many d ifferent materials , dimen·
sionless sensitivity coefficiell\s may not be practical. For our actual applications. we use sca led
sensitivity coefficients almost exclusively. The sensitivity coefficients for the volumetric heat
capacity and thennal conductivity are found 10 be

qL a¢
Tc = C -
aT
~ --u- Xc = -Tc- = - a¢
Ct - (14.8)
ac k au qL / k Ju

.T,~ kiik ~ qTL (a¢)

aT aaa -¢
T,
X,~ qL/ k ~aau-¢
u¢ (14.9)
446 SENSITIVlTY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

where
8¢ at
a-=- (14.10)
Ja U

Note that aT j ac and aT18k depend on their respective parameter values; this indicates (hal
the problem is non linear in C and k.
The foregoing three dimensionless sensitivity coefficienrs (X"' Xc. Xk) given by Eqs. (14.7)
- (14.9) have been computed and the results are given in Figs. 14.1 - 14.3. Notc that Xq is always
positive; increasing the heat flux increases the temperature. Xc is always negative: increasing
the volumetric heat capacity decreases the temperature. The behavior of Xk is more complicated.
Near the heated surface, the conduc tivi ty sensitivity coefficient is negative, wh ile the opposite is
true for positions ncar the back face. This means that there is a depth below the heated surface
for which the conductivity sensitivity coefficients arc small, indicating an insensitivity of the
temperature to the conductivity.
The magnitude of the sensitivity coefficient is also important. Since the dimensionless sen-
si tiv ity coefficicnts in Figs. 14.1~ 14.3 have becn scaled by qL / k (a characteristic temperature
rise) , their magnitudes are an indication of their relative importance. Since this example is
driven by the applied flux to one boundary, intuition is confirmed that the heat-flux sensitiv-
ity becomes smaller as you move away fro m the x :;::: 0 boundary . For si mple problems, many
analysts already have good intuition on the qualitative aspects (sign) of sensitivity coefficients.
However, the techniques presented in this chapter wi ll allow one to address the quantitative
aspects of sensitivity coefficients.

0.7

0.6
.VL=O.O

0.5
0.25

0.4
• 0.5
'" 0.3 0.75

1.0
0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5

FIGU RE 14. 1 Dimensionless heat-flux sensitivity coefficient for the {;onst:Ult heat-flux problem defined
by Eqs. (14.2) and (1 4 .3).
COMPUTATION OF SENSITlVITY COEFFICIENTS 447

1.0
-0.1 \.
0.75 0.25

-0.2
xlL = O.O

-0.3

-0.4

....().5 0
0.1 0.2

FIGURE 14.2 Dimensionless volumetric heat-capacity sensitivity coefficient for the constant heat-flux
problem defined by Eqs. (14.2) and (14.3).

1.0---1
~~:::-_ _ 0.75
0.1

..--c::07L---- -0.5 ----~

01LL=

0.25
-0. 1

-0.2
:elL:: 0.0

- 0.3

o 0.1 0.2 0.3 0.4 0.5

FIG URE 14.3 Dimensionless thermal conductivity sensitivity coefficient for the conStant heat-flux prob-
lem defined by Eqs. (14.2) and (14.3).
448 SENSITIVITY ANAL Y81S AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

One can analytically demonstrate that for a given (x, t) the foregoing three dimensionless
sensitivity coefficients sum to zero:

Xq + Xc + Xk = 0 ( 14.1 1)

This demonstrates that in this problem the sensitivity coefficients arc not independent of each
other. While this fact is important for parameter estimation (indicating only two of the three
parameters could be estimated from this experimental configuration), it is not of major concern
in the computation of sensitivity coefficients for design and uncertainty propagation purposes.
The primary use for differentiation of analytical solutions is to develop concepts and intuition
for simple problems and to usc analytical solutions to develop verification problems for the
numerical methods that are used for complex heat transfer problems. McMasters, et al. [61
USed Green's functions to develop analytical solutions to thousands of 3D problems for cubical
geometry with classical boundary conditions. These solutions could be analytically differentiated
to obtain analytical solutions to 3D sensitivity problems.

14,2,2 Finite-difference Methods

The term finite difference as USed here refers to the parameter space and not finite-difference
discretization algorithms discussed in earlier chapte~. It is one of the most important methods of
compuling sensitivity for complex heat transfer problems since it does not require modification
to the source code. The procedure is to run the computational model with nominal parameter
values p. A second run is made with a penurbed value for parameter Pi. A finite-difference
approximation in parameter space is then used to compute the scaled sensitivity coefficient from

_ ' Ch p, . .... Pi + ilpi.···. p,,) - , (fit. p,. .... Pi' ... , p,,) (_ )
'I'=~ +Oil~ ( 14.12)
I 6Pi
If there are II parameters, then n + I runs of the code will be required to compute the II fi13t-
order sensitivity coefficients. If a second-order central difference is used, then the number of
runs goes [0 2n + I. A second-order accurate finite difference is

r pi =Pi
_ '(P l' p,. ... . Pi + ilp" ... , p,, ) -, (fit. p, ..... Pi - ilPi.···. p,) + 0 ('_p',. )
Ll
26p;
( 14.13)
The primary difficulty with the finite -difference method is choosing an appropriate pcnurbation
size !:J.Pi ' If !j,:Pi is too large, the truncation error will be too large . If !:J.Pi is tOO small,
machine round-off becomes significant because of subtractive cancellation. Consequently, some
numerica1 experimentation is recommended. An example problem will help solidify some of the
issues associated with the fin ite-difference method (in parameter space) .
Consider the constant heat-flux problem for which the analytical was given by Eq. (14.4) .
This problem will be solved numerically using the finite-difference method in space and a first-
order fu lly implic it time integrator. The sensitivi ty coefficient will then be calculated using the
first-order finite difference in parameter space given by Eg. (14.12). The. problem parameters
are representative of a stainless steel and are as follows:

q = 4 x 105 W m- 2 k = 10 W m- 1K- 1 L = 0.01 m

1
p = 8000 kg m - J ell = 500 J kg-I K- Ti =300K (14.14)

at! = 05 at
_. --~ = 2.5
L2 . ilx-
COMPUTATION OF SENSITIVITY COEFFICIENTS 449

The problem was numerically solved using a first order in time and second order in space
finite-difference method; a first-order finite difference in parameter space was then used to
com pure the sensitivity coefficient h. The relative error in this numerical solution for Tk was
computed: the analyricaJ solution from Eq. (14.9) was taken as the exact answer. The computa-
tional domain is spatially discretized into unifoml clements; Fig. 14.4 presents the computational
results for 10, 20. and 40 elements. During the grid refinement, a!:::.t / !:::.x 2 was kept fixed; if
!:::.X was reduced by a factor of 2, then !:::.t was reduced by a factor of 4. Experience suggests
the following model for the error in Tt :

A(~k) (a~:) Cj (~k)

-----
e.Tk = + 8
~X·
+ (14.15)
--..-
parameter
discretization
-------
space/time
discretization
subtractive
cancellation

where A, B, and C are constants and I( . ) is an unknown function. If the space/time dis-
cretization errors were not present. then the results would be expected to follow the first-order
reference line shown in Fig. 14.4. As the grid is refined, the results approach the reference line
for the larger values of 11k / k. Further reductions in 6k/ k cause the space/time discreti7..ation
to become dominant and the relative errors reach a minimum plateau. As b.k / k is reduced
further, subtractive cancellation becomes dominant and the relative error can actually increase.
The range of !J.k / k for which the error in Tk reaches a stable minimum depends on the number
of clements. This range is broader for a coarse grid than for a fine grid . Hopefully, this example

10
,

~
'ff'

10 elements J
10'Ie c

.=
If
v
20 elemeflt~
£
'ic A
40 elements
/,, t order
ref line

V
, ,
10- 13 10- 11 10-7 10' 10'

FIGU RE 14.4 Relative error in finite-difference computation of Tk using a backward difference; numerical
discretization algorithm wa~ finite difference with a first-order implicil lime in!egralOf and the spaceltime
grid refinemenlmaimained ai:J.lji:J.x 2 = 2.5 .
450 SENSITIVITY ANAl YS\S AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

will provide some impetus to pcrfonn numerical experiments when using the finite-difference
method in parameter space.

14.2.3 Complex-step Method

T he complex-step method for computing sensitivity coefficients can be made independent of
the finite-difference step size and is a competitor to the finite -difference (in parameter space)
method provided source code is available. The method can be derived from a Taylor series
expansion of a rcal-valued function r about the complex (imaginary) parameter value p + i 6.p
(i = ..;=1). This series for a single parameter is

,(p+iflp) = ,(p)+ -aa' i iflp+ -211 "2,2 I

'
a', I
(iflp) , +31;;---'
-I '
(iflp) +0 (4)
flp
P p • fJp P • op I'

"' I 'flp
=, (p) + -iJp · -p
- 1 -0 ,
2!iJp2p
2
I ' 3!; vp0', Ip ' + 0 (')
flp - - - -
3
flp flp (14.16)

Taking the imaginary part of Eg. (14.16), one obtains

Im[dp +i flp)]= -0' I f l p - -1 (14.17)

Bp p 3!

Solving for the first derivative, one obtains

(14.18)

This very simple result indicatcs that the derivative of the function is obtained by evaluating
the function r at the complex parameter value p + i D.p. taking the imaginary pan, and dividing
by the step size D.p. Note that this result is second-order accurate. While it was claimed earlier
that the complex-step method is independent of step size D.p, it is clear from Eg. (14. 18) that
the result depends on step size. It will be demonstrated that the step size can be made arbitrari ly
small since the result does not suffer from the subtractive cancellation of the finite-difference
method. Consequently, from a practical consideration, a very small step size can be chosen and
one does not have to worry abour performing a parameter srudy to verify that the chosen value
was too small.
The same one-d imensional, transient, constant heat-flux example problem will be used to
demonstrate the complex-step method . We will use the same space/time discretization algorithm
(implicit, first order in time, sccond order in space) used in the Finite -difference Methods section.
The (real) input quantities to the code were converted to complex (as well as the tridiagonal
matrix direct solver) and Eq. 04. i 8) was applied to the complex temperature field. The thermal
conductivity sensitivity coefficient was the focus of this analysis and the relative finite-difference
step size was varied over the range 10- 15 .s D.kj k ::: I to demonstrate the broad latitude one
has in selecting lJ.k j k. The computational results are shown in Fig. 14.5. Experience suggests
that the error is of the form.

CT, = A(~k)' + R(afl')

flx'
(14.19)

--------
parameter
discretization
'-.-------'
space/time
discretization
COMPUTATION OF SENSITIVITY COEFFICIENTS 451

10
,

10c1cmcnts
./
g 100 10- 1 c
"~
·0

"
""
10 '
~';'~':'~':'~':'
20

40
r
1 p= 2

10-13 10- 11 II,..' 10- 1

6kJk

FIGUR E 14.5 Relatille error in T* as a function of step size llk/k for 10.20, and 40 elements: the
space/time grid refinement maintained alll / tJ.x 2 ::: 2.5.

Note the absence of the subtractive cancellation in the error term and that the parameter
discretization is second order. For the parameters considered. any 6k/ k !: 10- 3 is adequate
to make the parameter discretization error smaller than the space/time discretization error. As
elements are added, the results become closer to the p = 2 line and the permissible upper bound
on t::.k / k becomes smaller. Comrast these results with finite differences (in parameter space).
as presented in the previou s section. The only penalty with {his method is implementation time
and increased CPU time due to complex arithmetic. Additional infomlation on the complex-step
method can be found on the web site [71 and Martin s et a1. [8] . Aerodynamic application s of
[he complex-step method can be found in references [9] and [10].
In contrast. the finite -difference method starts with n small trial value for t::.k followed by
check to see if the results arc sensitive to 6k, wh ile the complex-step method simply uses a
very small value of 6k.

14.2.4 Software-differentiation Me thod

The software-differentiation method is a computer science development. An existing source code
(FORTRAN 77 or C) is input into a special preprocessor (ADIFOR or ADIC for FORTRAN
77 and C. respectively) that performs line-by-line differentiation of the original source code
while producing a new source code for the sensitivity coefficient. Examples of this technology
are presented by Bischof et al. III J. where they have successfully applied it to large codes. If
there are multiple parameters for which sensitivity is desired, multiple runs of the preprocessor
ADIFOR or ADIC are required. If the original source code is modified (enhancements, bug fixe s.
etc .). then the preprocessor must be run again. Since our work has been focused primarily on
techniques that can be readily applied to software under development as well as many parameter
problems, we have not personally exerci sed the software-differentiation method s.
452 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

14.2.5 Sensitivity Equation Method

From the continuum equation perspective, the sensitivity coefficient is a field variable just like
the temperature or ve locity field. This section presents the sensitivity equation method (SEM) in
which field equations describing [he sensitivity coefficient will be developed. T he SEM involves
differentiating the partial differential field equations with respect [0 [he parameter(s) of interest.
This process will yield new panial differential equations for the sensitivity coefficients. These
new equations arc then solved usi.ng the same numerical techniques as were used for solving
the basic field equations.
To demonstrate the SEM, consider a one-dimensional planar slab of thickness L w ith a
specified flux q£ plus a far-field radiation boundary condition on the left face (E, Tr) and con-
vection on the right face (h. 1 ~) . These boundary conditions were chosen to provide a variety
of boundary conditions, including one that is nonlinear. The equations and boundary conditions
for this model problem arc given as follows:

aT aq
c-+-~, o ( 14.20)
at ax
aT
q ::: - k - (14.21 )
ax
(14.22)

q l ., _,~, - k -aT I ~ h (T - 1;")1, =, (14.23)

ax ;r- L

T(x.O)=7i (14.24)

The parameter vector for this problem contains eight clements and is

(14.25)

The fact that this extremely simple heat transfer problem contains eight parameteI1i indicates
that prac£ical heat transfer problems can easily have IDs to IODs of parameters .
We must now differentiate Eqs. (14.20) - (14.24) with respect to each element of the parameter
vector. Starting with the volumetric heat-capacity sensitivity coefficient, we obtain

ac
(aT aq )
-a c -+--
ar ax
~c -
a (OT) aT a (Oq)
at -ac +-+- -
ar ax 11 C
~ O (14.26)

where it has been assu med that the order of differentiation can be interchanged. If the above
equation is multiplied by the nom inal value of C. the scaled sensitivity coefficient is naturally
introduced. The result is
c -a
ar
(at) + c -ar+ -a ( c--'!.aca ) ~O
C-
dC
aT
dX
(14.27)

The sensitivity of the local heat flux to C is found by d ifferentiating Eq. ( 14.21) to yield

(14.28)
COMPUTATION OF SENSITIVITY COEFFICIENTS 453

Equation (14.27) can now be written as

aTc- - - k -
C a ( aTe) ~ -C-
aT (14.29)
at ax ax at

Equation (14.29) is the partial differential equation that describes the field variable Te. Note
that the left-hand side is identical to the left-hand side of the original energy equation; this is a
common characteristic of the SEM. However, the Tc equation has a right-hand side term that can
be viewed as an apparent source tenn. If the temperature field is computed first, then this apparent
source tenn is simply a known function (of time/space). Your favorite numerical algorithm for
determining the temperature fie ld can also be used to solve the sensitivity field equation.
We will continue the development of the Tc equations by differentiating the initiallboundary
conditions with respect to C. From Eqs . (14.22) and (14 .28) we obtain

1c l... =o = aq l
e 'Ie = -k aTe l
-a- = - 4&0 (T 3 Ie) I
;r-() (14.30)
() .10=0 X .10=0

While the left-hand side (x = 0) boundary condition for the energy equation was nonlinear and
inhomogeneous. the corresponding Tc boundary condition is linear and homogeneous. Thi s
assumes again (hat the temperature field is known prior to the computation of the sensitivity
field. Through a similar procedure, the boundary condition on the right face is given by

a I
e-.!!... aT
= - k --.£ I
= hIc lx=L (\4.3\)
ac x=L ax x;L

This boundary condition is linear and homogeneous. Since the initial condition is indepen-
dent of the volumetric heat capacity, the corresponding initial condition for Tc becomes the
homogeneous condition
aTI
C- Tcl (.T .O) = 0 (14.32)
ae (x. O)

The formulation of the field equation and associated boundary/initial conditions for Tc is com-
plete and is given by Eqs. (14.29) - (14.32). Due to the similarities in fonn of the T and Tc
equations, the same technique can be used to numerically solve these equations. It docs not
matter if the discretization algorithm is finite difference. finite volume. or finite element. In fact.
the coding for discretizing diffusion and capacitance terms for the T equation can be used for
the similar tenns in the Tc equation.
The computational procedure is to time march the discretized equations one time step and
compute the T field at the new time. From the temperature field , the right-hand side of the Tc
equation can be computed along with any boundary-condition tenus. The Tc equation can be
integrated one time step . The solution for the sensitivity field is segregated from the solut ion for
the temperature field. Even though the T equation is nonlinear, the Tc equation is linear. This
linearity is one of the significant computational advantages of the SEM over the finite-difference
method. For a nonlinear PDE, each perturbed parameter vector will require a nonlinear solution.
By inspection of Eq . (14.29), it becomes obvious that Tc is controlled by the time rate of
change of the temperature field. If the temperature field approaches steady state. Tc should
approach zero. While this is obvious without a discussion of sensitivity coefficients, insight
into more complicated problems can be gained from studying senSitivity coefficient describ-
ing equations.
454 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

The 1;' equation is morc complicated than the 1(: equation since k appears in both the PDE
and the boundary conditions. Following the same procedure as above. the differential cqu<ltion
for T" can be written as

( 14.33)

Differentiating Fourier's law with respect to k yields

Jq ar,
q, =k- = - k - - k -
aT (14.34)
[)k ax ax
While Fourier's law involves a single term, the sensit ivity of q to k involves two terms. The
first term involves what can be thought of as a flux of sensitivity information plus a second
term that is the heat flux itself. Combining Eqs. (14.33) and (14.34), the 1;' equation becomes

(14.35)

Again, the left-hand side of the Tk equation is identical in form. to the T equation ; the right-hand
side has a fictitious source tenn that is the negative of the gradient of the heat flux. Gradients
of the local heat-flux drive the Tk field.
Care must be exercised in deriving the boundary conditions for the Ii equation. Intuitively.
one might be inclined to derive a boundary condition on kaTk!ax. However. we need a condition
on qk that can be derived by differentiation of the right-hand sides of Eqs.(14.22) and (14.23)
with respect to k. The results are

( 14.36)

( 14.37)

Again. the nonlinear, inhomogeneous boundary condition for the T equation has become a linear.
homogeneous boundary condition for the Tk equation . The initial condition for 1k is simply

(14.38)

We have addressed two of the three gradient-type boundary conditions that commonly occur.
The third type of gradient boundary condition is a specified heat flux qs. Since the magnitude
of a specified flux q$ is independent of either k or C. these boundary conditions become

(14.39)

where the subscript b designa£es the generic (left or right) boundary along which this boundary
condition is applied. Note that for Tko JTk/ax of:. 0 along a specified flux boundary. This subtle
point r\!quires careful thought.
COMPUTATION OF SENSITIVITY COEFFICIENTS 455

The specified temperature is the only remaining boundary condition we have nor discussed.
Since a specified temperature is independent of either k or C, specified temperature boundary
conditions translate to

Ie IXh = Ii IXh = 0, along specified T boundaries (14.40)

Of the parameters listed in Eq. (14.25). k and C are spec ial in that they both appear in the
describing POE. For their respective sensitivity-describing equation, an inhomogeneous term is
prescnt. For all other parameters thaI do not appear in the T equation, their sensitivity-describing
equation can be written as

P '" k or C (14.4 1)

where p is a generic parameter.

The boundary conditions for the remaining parameters in Eq. (14.25) arc now discussed. Dif-
ferentiating the x = 0 boundary condition given by Eg. ( 14.22) with respect to these parameters
results in
f°'ael l
x=o
=f[O(T:-T')-4<oT ,
JT
ae
llx""o (14.42)

= [<0 (T: - T4) - 4<oT 3 T, 11.; o (14.43)

(14 .44)

(14.45)

(14.46)

(14.47)

Rather than using I~ and 1'00 to scale their respective sensitivity coefficients, a reference tem-
perature change l:J.I ~ or 1:17'00 is used. This eliminates problems with lero temperature when
absolute temperature units are not used. These reference temperature changes represent a char-
acteristic temperature change for the problem. As an example, one might choose the maximum
temperature rise of the system, l'max - Ii. The same reference temperature risc could be used
for both Tr and Too sensitivities. although this is not necessary. Since the deSCribing equation
is homogeneous for those parameters that do not appear in the energy equation, the inhomo-
geneities in the boundary or initial condit ions will drive the remaining sensitivities. For example,
the f: sensitivity in Eq . (14.43) is driven by the radiative heat-flux term sa (1/ - 1'4).
By now, one should see a pattern developing in the sensitiv ity equations . With this in mind,
the remaining results will be given as

(14.48)

(14.49)
456 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

oq
h --
oil
I
x= l.
( 14.50)

D. Toc ~~
aT I = h (T1", - D. Toc ) Ix= L (14.51 )
00 .f""L

The inhomogeneities in the hand 7'00 sensitivity boundary conditions at x = L arc the convective
heat fluxes 11 (T - 1'x) and htJ.TCXh respectively.
We have discussed the initial condi tions for both Tc and Tk. Eq s. (14.32) and (14.38) ,
respectively, and they are both zero. It is easy to see that if the parameter of interest is anything
other than the initial temperature itself, the initial co ndition for I" will be zero. The initial
conditions can be summarized as fo llows:

p, !pT
J
I
(x.O)
= Tp, 1'<.0) = I 0, Pi :l= 1;·
tJ. 1;·, Pi = I i
(14.52)

As with other sensitivity coefficients related to temperature, we have used a temperature change
as a scale factor.
We now summarize all of the results developed in this section in tabular form. First, we
present the boundary condition types in Table 14.1. The ~x .~ term allows a given boundary
condition type to be applied on either the left- or righ t-hand face. All of the one-dimensional
results are summarized in Table 14.2. The way the boundary conditions arc written allows all
possible pairs of boundary conditions to occur; the same type of boundary condition could be
applied to both surfaces. The numerical solution of each sensitivity equation can be implemenred
using your discretization method of choice.
After implementation of the SEM. the first step is to perform verification calculat ions to
insure that the implemented equations arc being solved correctly. The finite difference code
used to compute the results presented in Section 14.2.2 was modified to usc the SEM method
and the results arc presented in Fig. 14.6. The discretization error in the method is first order in
time and second order in space. Consequently, as the grid was spatially refined, we maintained
at:J.r / t:J.x 2 fixed. The exact value for 'Ii was computed from Eq. (14.9) and the relative error
was computed from

re 1atlve . /.~:::::::: 100 -,--,k

. error III I
T~,"~,~,:-_k
- 7:~"~"m", I (14.53)
Tk;rnol

For comparison purposes, a line with a slope of -2 is shown. Clearly. the errors are asymptotic
to the second-order line. This gives us confidence that the method was implemented properly.

TABLE 14.1 Classification of Four Common Boundary.condition Types

Be Type Name Boundary Condilion

Isotherm.al
2 Heat flu:~

3 Convection
4 Radiation
COMPUTATION OF SENSITIVITY COEFFICIENTS 457

TABLE 14.2 Summary of Right-hand Side, Initial Conditions, and Boundary Conditions for
Sensltivity-eoeflici ent Equations

qPi L<~ q p, I ~~ qp, Ix~

T" RHS Ie Type 1 Type 2 Type 3 Type 4

'lie ..£..
~x
(.I;: !r)
~, <
0 0 0 II J( IXl 4euT3 Tl. I ~.

Tc - c ~; 0 0 0 II Tcl~1 4
4euT3 J(' IX
J~ , 0 0 0 II T.. I.'J 4euT~ Tt4
J~ 0 0 0 q II 7;,1 4 r.uT Tq x~
T, 0 0 0 0 "
II Thl .\., +{h 4eaT.l Thlx~
T, 0 0 0 0 h J ~lxJ 4caT-' ·' ~lx1 + qr
TT, 0 0 0 0 flUT"" I.\ 3 -I:J. T00, ) 4eaT] TToc .,
Trr 0 0 0 0 II T·r.1 4 r.a( T 3 TTrl - TrJ I:J.Tr)
Tr. 0 0 f1T, 0 II
,
TThl
" 4eu r
J "
TTbl •.
TT, 0 I:J.T; 0 0 II "
Tr, I.n 4ca T J fr, Ix:

The SEM results are consistent with the finite difference in parameter space results given in
Fig. 14 .4. without having to iteratively choose tI value for !:!.k / k. However, th is was at the
expense of having to modify the sourCe code. There is obviously a trade-off between code
development ti me and time required to iteratively dctennine thc appropriatc value for D.k / k for
your specifi c problem.

x =O

p =-2

10- 1

to 20 30 40
Numher of clements

FI GURE 14.6 Grid-fefi nemell1 error in Tk using SEM for 10 conslalll q problem; x = 0 for t = 4 s
(ut / L 2 = 0,1) and 20 S (ul / L 2 -= 0.5) with al:J.l / I:J.x 2 = 2.5 for all calculations.
458 SENSITlV!TY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

14.2.6 Sensitivity Equation Method for Multidimensional Problem s

All of the results presented so far have been for one-dimensional problems. For the SEM to be
of prac tical value, it must be readily extendable to multidimensional problems. The results given
in Tables 14.1 and 14.2 can be written for three-dimensional problems by using the following
substitutions:

-k- aT (A A)
e~·n ~ - (k . "VT).f!A (k"VT) A
.11 ( 14.54)
ax '-.-' 0::; -

~
i!;Otmpic

-a ( k--
aT) ~V.(k . VT)~V . (kVT) (14.55)
ax ax '-..-' '-,--'
31liSQIrOpic

Blackwell et at. [121 and Dowding et al. [131 have implemented the SEM for heat conduction
in a multidimensional control-volume finite-clement code. Results from this code have been
Ulilized to compUle sensitivity coefficients for a thermally activated battery. as discussed in
Blackwell et al. [14] and have been used for a number of other practical hcar transfer problems.
Pelletier, Borggaard, and colle::lgues have ::lpplied the SEM for a wide variety of ftuid-
flow and heat transfer problems . For example. [15J considered nonl inear heat conduction and
[16J considered second-order sensItiv ity analysis for laminar flow; these citations wi ll lead to
nUmerous other works by these researchers. In the mechanics arena, see Kleiber et al. [17. 18].

14.2.7 Differentiation of Discrete Equations and Adjoint Methods for Steady

Conduction Problems

In instances where sensitivity infonnation is not required at all locations in the computational
domain, the adjoint method may be computationally more efficient than the SEM. To demonstrate
this, consider the discretiLed equation for steady heat conduction

K T=s (14.56)

where K is the global conduction matrix. T is the vector of unknown temperatures , and s is the
right-hand-side vector. Differentiating this equ:.ltion with respect to one element of the parameter
vector yields
aT il K as
K - - c- T ~- i=I. ... . ll p
ap; api api

Introducing the scaled sensitivity coefficient allows one to write

as il K
i 1.... , n p (14.57)
0::; Pi-.- - Pi- T 0::;

api fJPi

For each Pi , :.In additionill system of linear equations must be solved. For linear problems, this
approach will be identical to the SEM provided the original field equation(s) and the correspond-
ing sensitivity equations are discrctiLcd consistently, For an clement assembly methodology as
in control-volume finite element or Galerkin finite element, the element-level equations can be
differentiated with respect to parameters of interest and then assembled into the global matrices.
COMPUTATION OF SENSITIVITY COEFFICIENTS 459

In parameter estimation work. the sensitivity coefficients may be desired only at sensor loca-
tions. For example. in the estimation of thermal properties from temperature measurements,
a finite number of sensors are used and the sensitivity coefficients are desired only at thcsc
locations. Following Kirsch [19]. thc adjoint method can bc devcloped by multiplying thc
sensitivity-coefficient equation, Eq. (14.57), by the inverse of the global conduction matrix to
yield
K - 1K T A = K - 1 r ~~
vp;
(p
I ~
VPi
_p."KT) (14 .58)

The left-hand side of Eq. (14.58) yields the vector of sens itivity coefficients for parameter Pi
at all nodal locations; however, we are concerned only with the sensitivity coefficient at certain
selectcd locations. To cxtract the sensitivity coefficient at a single location, define a row vector
that has zeros everywhere except for unity at the jth nodal location

I; = (O ... 0 I 0 OJ (14.59)

Multiplying Eq. (14.58) by Eq. (14.59) yields

T () T
I jT,,; :;: : : T"i j :;: : : l j K
_I (as
Pi
aPi
aK )
-PiaPiT (14.60)

Equ<ltion (14.60) gives the sensitivity coefficient for Pi at nodal loc<ltion j. It is convenient to
define the coefficient of the parentheses of Eq . (14.60) as the adjoint variable vector, which is

(14.6 1)

Taking the transpose yields

(14.62)

which can be written as

(14.63)

Although Eg . (14.63) is valid at allllrnodallocations, the adjoint variable approach is attractive

only when the number of sensors li s is a small subset of thc number of nodes II. Note that Eq.
(14.63) is independent of the panicular sensitivity coefficient one is trying to compute; this
means that the ndjoint v<lriable vector depends only on the spatial (sensor) location and K. Once
Eq. (14.63) has been solved for the adjoint variable vector ~j, the sensitivity coefficient for
all parameters of interest (at this noon I location) can be computed from Eq. ( 14.60), which is
written as

i = I, .. .. II P' j = I .... , li s (14.64)

Equation (14.57) defines the discrete fonn of the SEM, while Eqs . (14.63) and (14.64) define
the discrete adjoint equations. Both approaches have a single left-hand-side matrix but multiple
460 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

right-hand-side vectors. The number of right-hand-side vectors can be used as a rule of thumb
for when one method is computationally more efficient than the other:

Usc SEM when Ilf> < 1Z s ·

Usc adjoint when 11" > It s_

Obviously, when IIp and li s are approx imately equal. this rule of thumb will have to be inspected
more closely.
For discretization methods that produce a symmetric K matrix (c.g .. Galcrkin finite element).
the left-hand side of the adjoint equation is identical to the original discretized field equation.
Consequently, the adjoint method may be more efficient for Galcrkin fin ite-element methods
(han for Olher disc retization schemes. This is true only if one is able 10 store the K matrix. For
the million unknown class problem, iterative methods wil l likely be used for the linear solver
and the K matrix will never be stored explicitly. As with the SEM, the adjoint method needs
coding to compute iJ S/iJPi and aK/ Jpi '

14.3 FIRST·ORDER PROPAGATION OF UNCERTAINTY IN COMPUTATIONAL

MODELS

Several techniques have been presented for computing sensitivity coefficients. We now want to
address the issue of uncertainty in model input paramete rs and their impact on the results from thc
model. If all parameters are known with the same precision, then the most important parameter
will be the one with the largest sensitivity coefficient. In practice, this is not the case. Some
parameters will be known from well-controlled experiments, while others are simply engineering
estimates from a variety of sources. The computational model with uncertain parameters is very
much analogous to an experimental data reduction equation containing uncertain paramcters. The
uncertainty propagation equation f(lr result r (x, t, p) is identical to that used by the experimental
community and is the sum of squares expression

(14.65)

where apJ pi is the relative uncertainty in Pi and r pi = Piar / Pi is the scaled sensitivity
coefficient. Some authors refer to this method as the mean value method, as only mean values
of parameters are required. For thl! deve lopment of this equation, see Coleman and Steele flO].
If the scaled sensitivity coefficients are known at every point in the computational domain, then
the uncertainty estimation is simply a postprocessing of these results.
An industrial example of uncertainty propagation through a computational model of a thermal
banery with 19 uncertain par<lmeters (thermal properties and convective heat tr<lnsfer coefficients)
is contained in Bbckwell et <II. [14]. Through the uncertainty analysis, it was found that only
four of these parameters contributed significantly to the overall uncertai nty. This me<lnt that
engineering estim<ltes were adequate for 15 of the 19 parameters. If the overall uncertainty in
the result r was too large for the system requirements, then resources could be directed toward
reducing the uncert<linty in the four significant p<lrameters.
Equation (14.65) is consistent with a linear-in-the-p<lrmneters model. Consequently, if the
model is nonlinear, then the estimated variance will be in error. Higher-order models arc dis-
cussed in Fadale [2 1J and Fadale and Emery [22]. It is our op inion that the uncertainty in
FIRST-ORDER PROPAGATION OF UNCERTAINlY IN COMPUTATIONAL MODELS 461

the uncertainty of the parameters is often sufficiently high that it is difficult to justify the addi-
tional computational/analyst expense of a higher-order analysis. Also, higher-order methods may
require knowledge of probability distributions, which are difficult/costly to obtain for practical
engineering problems. The uncertainty calculation should bc viewed as an estimate as opposed
to a precisc result since many assumptions arc necessary to perform an uncertainty analysis. The
most important result from an uncertainty analys is is understanding which parameters dominate
the overall uncertainty estimate. This naturally leads to a discussion of importancc factors, which
give a relative ranking of the various parameters in tenns of their contribut ion to the overall
uncertainty. If Eg. (14.65) is divided through by the left-hand side, we obtain

0' ( , )' + -0" (',,- )' + ...

1 ap ! I a p2
1~ - ,', - (14.66)
, pI r p2

The importance factor for Pi is simply

a )'
importance factor for Pi = I Fi = 21 ( r Pi ...J!J.. (14.67)
a, Pi

Alternative methods for computing importance factors are given in Helton and Davis [23] .
The uncertainty in the predicted temperature will be computed fo r the one-dimensional con-
stant heat-flux example with properties given by Eg. ( 14.14). The uncertain parameters are q, k,
and C with a relative uncertainty of 0.05 for each parameter. Figure 14.7 presents the estimated

600 2aT error bars

,Ipi::: O.05
0 1'

T
xlL::: 0
1

400 T
xlL = I
.L

HGURE 14.7 Estimated uncertainty in lIlodel temperature lxj L:: 0,1 ) due to uncertainty in q, k, and
C: all relative uncertainties are 0.05. The mean value method was ll';;ed, Eq. (14.65).
462 SENSITIVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

uncertainty in temperature at the fron t and back surface; the analytical sensitivity coeffic ients
were used. Due to the (estimated) parametric uncertainty, the front face temperature may vary
by as much as ±40 K. This ±40 K is the range that the model output could span given the
uncertainty in the inputs. The reason for labeling these results as 20 is that the values assumed
for the parametric uncertainty were 20 values; i.e., 95% of the rime the parameter values wi ll
lie in the ±2a range (for normally distributed parameters). The importance factors defined by
Eq. (14.67) were also calculated but they are discussed in a subsequent section.

14.4 SAMPLING METHODS FOR SENSITIVITY AND UNCERTAINTY ANALYSIS

T he methods presented In the preceding section could be rcnned local sensitivity and uncertainty
methods because the parameter space is sampled only in the neighborhood of nomin<J] parameter
values. If the response for off-nominal parameter!; is quite different from that for nominal
parameters, then sampling-based methods should be considered. However. the design will be
nonrobust. if for a dominant parameter(s), the nominal and off- nominal responses are quite
different. For high-consequence events, one might question the advisability of having nonrobust
designs (at least for the dominant parameters) unless the dominant parameters arc known with
a high degree of precision.
A global uncertainty method is to sample the parameter space using Monte Carlo methods.
T he modern versions of Monte Carlo methods trace their roots to the nuclear reactor safety
community. A full Monte Carlo analysis might require 100s- 1000s of function evaluatiOns to
insure convergence; for the million unknown class of problem, this approach is not computation-
ally feasible in today' s computing environment An alternative is the Latin hypercube sampling
(LHS) method developed by Mc Kay et al. [24]. In the LHS method, the cumulative probability
distribution of a given variable is divided into lILl1S (2:: 11" + 1, generall y when sensitivity is
desired) bands of equal probability. Within each band. a random sample is drawn. This process
is repeated for each of the 11" model parameters (or variables) such as thermal conductivity, heat
capacity, and volumetric source. The matrix of IILHS x lip values is represented in Table 14.3.
The columns in Table 14.3 represent the LHS samples for a given variable, while the rows
represent the model parameter vector for a given probability band. To insure full coverage,
the model parameters are combined in a random fashion in a process described by Helton and
Davis f2, 25] as follows: " The Il LHS values thus obtained for PI are paired at random and
without replacement with the llLHS values obtained for Pl. These IlLIlS pairs are combined in
a random manner without replacement with the lILHS values of P3 to form llLHS triples. This
process is continued until a set of lILHS lip-tuples is formed." The above methodology has been
documented in Iman and Shonencarier [26 ) and Wyss and Jorgensen [27) and is available in

TABLE 14.3 Matri" Representation of Number of LHS

samples ( ULHS) and Number of Parameters (u p)

Parameters
Probability Band p, 1', p",
I p" p" Pl" p
2 p" P22 "'",
IILHS J)"Ulsl P"1.Hs2 P"UlS" P
SAMPLING METHODS FOR SENSITIVITY AND UNCERTAINTY ANALYSIS 463

the dakota software [28]. The LH S approach should better capture nonlinear beh:::lvior over the
sampled parameter space.
Once the computational model has been run for the IILHS parameter vectors, standard statis-
tic:::ll techniques can be used to process the results. Estimates of the expected value and variance
of response r are given by

_ I "LHS

r= E (r) = -
IlLHS ;= 1
L: r; ( 14.68)

I /lLHS

VV) = ------,-
flLllS - 1
L: h - EV)l' (14.69)
j: J

where V (r) = o}. If the expected response and its variance arc the only things of interest,
then the computational process is complete. However, it is generally desirable to rank order
the pammeters in tenns of their signific:::lnce. This can be accomplished using the techniques
outlined in the following p:::lmgraphs.
O ne technique for determining the global sensitivity is through a regression analysis of the
respome r as a function of the parameter values Pi. The simplest approach is to assume a linear
relationship between response and parameters of the form

"r
r = ha + Lbj p j (14.70)
j= 1

where the b/s are regression coefficients: if the parameters are correlated. then a mOre sophis-
ticated model would be required. The sensitivity of the response to changes in the model
parameters can be obtained by differentiating Eq. (14.70) with respect to the parameter of
interest. yielding

or'
- = hi (14.71)
api

This tifS[-order (in pammeters) surrogate or response-surface model of the L HS results gives
global sensitivities that are analogous to the local sensitivity coefficients obtained using finite
differences. Using the sensitivity coefficients computed from Eq. (14.71), the importance factors
can be computed from Eq. (14.67).
Standard techniques can be used to compute the regression coeffi cients in Eq. (14.70). How-
ever, a word of caution is appropriate. Since the sensitivity coefficients have units associated
wi th them, they may vary by orders of magnitude. For example. the volumetric heat capacity and
thermal conductivity of 304 stainless steel at room temperature arc approximately 3.7 x 106 J
m- 3 K- l and 14.5 W m- J K- l
, respectively. This suggests that the regression equations should
be written in the fonn

(14.72)

and {3j is solved for directly. Farone example problem. this approach reduced the condition num-
bcr of the linear regression equations from 4 x 10 15 to 8.9 x 103 . Another approach to addressing
the conditioning problem is tbe singular value decomposition method: see Strang [29] for details.
464 SENSlTlVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

Once the scaled sensitivity coefficients arc detennined from the linear-regression 3m11ysis, the
importance f<lctors can be calculated fro m Eq . (14.67). A higher-order regression analysis can be
performed in conjunction with the LHS method, but additional LH S samples may be required.
With sampling-based methods, there may be some question if the number of samples was
adequate. O ne way of answering this question is to pClfonn replicates. With the LH S procedure.
this is accomplished by starti ng the random-number ge nerator with a different seed. The entire
analysis is repeated and the results arc compared for the different replicates. T he replicates can
be processed individually or as a group.
The above LHS methodology has been app lied to the COnstant heat-flux prOblem used
throughout this chapter. Ten LHS runs were made with the finite-difference nu merical code;
the model contained II equally spaced nodes. The three variables {q k C) were assumed to
have a log-normal distribution, eac.h with a relative standard deviation of 0 .05 ; the correspond-
ing LHS scale factors arc given in Table 14.4. Equations (14.68) and (14.69) were used to
compute the average and standard deviation of the nodal temperatures; the results for x / L = 0
and I are shown in Fig. 14.8. For comparison purposes, the complex-step method was used for
the sensitivity coeffi cients in conjunction with Eq. (14.65) to estimate the standard deviation in
the temperature . T he results at x/I. = I compare very favorably but those at x/L = 0 less so .
Two reasons for the discrepancy are as follows : first. the error propagation given by Eq. (14.65)
is an <lpproximation; and, second. 10 LHS samples may not be adequate for this problem. These
results point out the need for perfonning grid-refinement studies to force the grid discretization
errOrs to an <leceptable level and to insure that converged LH S results are used. However. user's
of this technology should keep in mind that all these techniques require statistical infonnation
about error in the parameters which, in many instances. arc simply engineering assumptions.
The linear response surface method described above was used with the 10 LHS runs (FD code.
II nodes) to compute scaled sensitivity coefficients and these results arc shown in Fig. 14.9. For
comparison purposes the complex-step sensitivity coefficients are also shown . T he two methods
for computing the sensitivity coeffi ,;ients (both using finite-difference discretization on the same
grid) agree quite well.
The parameter importance factors. as defined by Eq. (14.67). have also been computed for the
constant heat-flux example. The importance factors wiJl allow one to rank order the parameters
in terms of their contribution to the variance in the predicted temperature field. Again. two
methods were used: the complex-step method for sensitivity coefficients in conjunction with
Eg. (14.65) for the propagation of variance: and LHS in conjunction with a linear response
surface for sensitivity COefficients. Results fOr the twO methods are shown in Fig. 14.10 and
the results are consistent. One should nOt focus too m uch attention on the magnitude of the

TAB LE 14.4 LHS Multipliers for the Three Parameters q, k, and C

Sample q k C

9.45944 X 10- 1 9.69835 X 10- 1 9.57020 x 10- 1

2 1.01863 x 100 9.45728 X 10- 1 1. 0678 8 x 10°
3 1.09567 ;.: 100 9.86180 x 10- 1 1.02313 x 10°
4 9.21243 X 10- 1 1.05484 x 10° 1.04920 x 10°
5 9.98533 X 10- 1 1.03684 x 10° 1.00529 x 1(jJ
6 9.65651 ~: 10- 1 9.87945 x 10- 1 9.87167 X 10- 1
7 1.00834 ~: 100 8.88581 x 10- 1 9.74585 X 10-
1
8 9.79963 ~: 10- 1 9.99364 x 10 I 9.62540 X 10- 1
9 1.03053 ~: 1(jJ 1.11242 x 100 1.03458 x lrfl
]0 1.04461 x I (jJ 1.02519 x lQO 9.08745 X 10-
1
SAMPLING METHODS FOR SENSITIVITY AND UNCERTAINTY ANALYSIS 465

50

45
- - mean value, FD with complex step
40
••••..•. 1-<1), LH S
J5

30

'•"c "
N

20

15
/
JO xlL = I

0
5 10 15 20
Time, ~

FIGURE 14.8 Standard deviation in temperature at xlL := 0 and I for constant heat-flux problem using
\0 LHS runs and complex step. The runs were made with finite-difference (II-node) numerical code.

350

300
250
200
'E" ISO
.~

- u
~

8u 100 - - mean value, FD with complex step

.," 50 FD,LHS
.".;;; 0
"•u
1l - 50
" - 100
u
~

-1 50

-200

-250
0 5 JO 15 20
Time. s

FIGURE 14.9 Temperature sensitivity coefficients at xlL:= 0 for constant heat-flux problem using com·
plex step and LHS with linear response surface model. The runs were made with finite-difference (II -node)
numerical code. Note the close agreement between the two methods.
466 SENSIT IVITY ANALYSIS AND UNCERTAINTY PROPAGATION OF COMPUTATIONAL MODELS

0.7
-----------------q----------~

06 -
- - - mean value, FD with complex step
LHS
0.5

!lw x/L=O
0.4 -
"•"
w
0

~
0.3

0.2 - ,--- C

0.1 ------ k

0 I I
IU
I
15
3
0 5 20
Time,s

=
FIGURE 14. 10 Comparison of importance factors for COIlSlalll flux example (xl L 0) as obtained from
complex step and L HS with FD (II-node) solution. Note the close agreement between the two methods.

differences in the two methods but instead should focus On the fact that the rank ordering is the
same for both methods. If one wams to reduce the standard deviation in the model predictions,
then reductions in the errors in the heat flux will be much more fruitful than reductions in the
errors in the other two parameters, Insight into the dominant parameters is the most important
outcome from a computational uncertainty analysis.

14.5 SUMMARY

Several methods have been presented for computing sensitivity information. While each method
has its strengths and weaknesses, there is no single best method for all applications. For the
engineering problems we have solved, at least two methods generally get used. A computa-
[ional uncertainty propagation through a model requires a lot of assumptions about parametric
uncertainty. In many cases. these assumptions may be no better than engineering estimates. Con·
sequently, the uncertainty propagation process should be viewed as a means to understanding
which parameters control the simulation results. The important thing is what is learned from
the process and not a precise statement of the computational uncertainty. The learning process
can help develop a road map On how to beSt spend project resourCeS to reduce the estimated
computational uncertainty. A quota.tion by Dowdell [30] summarizes these scntiments: "T here
is a lot of uncertainty on uncertainty analysis but some analysis is better than no analysis at all:'

ACKNOWLEDGMENTS
This work was funded by Sandia National Laboratories, a multiprogram laboratory operated by Sandia
Corporation, a Lockheed Martin Company, for the United Slates Department of Energy under Contract
DE-AC04·94ALR5000.
NOMENCLATURE 467

NOMENCLATURE

A.H.C constants
bo.bj regression coefficients
cp specific heat
C volumetric heat capacity
E(. ) expected value
e volumetric source
e" error in P. approxi mate-exact
unit vectOr in x directiOn
"
f( · ) function

"
I
convection coefficie nt
identity matrix
Ij identity vector
IF,' impOrtance factor
; imaginary number cJ=1)
K global conduction matrix
k thennal conductivity
L length
n number of nodes
IILHS number of LHS samples
np number of parameters
11 ,< number of sensors
n, number of times
,; unit vectOr in nonna l direction
P, Pi parameter vector, parameter element
q , qQ, qs heat flux
qp scaled heat-flux sensitivity coeffi cicnt
,. response function
, mcan value of r
s right-hand side vector
T temperature
T, initial temperature
1~ far-field radiation temperature
1'00 cOnVectiOn temperature
Tp scaled temperature sensitivity coefficient
I time
If final time
11(. ) variance
Xp dimensionless temperature sensitivity coeffic ient
x,x position vector. position

Greek symbols

a thermal diffusivity
61 time discretization
6x space discretization
6p parameter perturbation
e emissivity
468 SENSITIVITY ANAlYSI8 AND UNCERTAINTY PROPAGATION OF COMPUTAT IONAL MODELS

, adjoint variable vector

viscosi ty
"
p
a
density
Stefan Boltlmann constant
response uncertainty (standard devi<ltion)
parameter uncertainty (standard dev iation)
temperature uncertainty (standard deviation)
analytical solution for constant fl ux
mean value

REFERENCES
I. D. G. Gacuci, Sensitivity ami Ulicenailll), AnalySiS. Vol. I; nleory, eRe Press, Boca Rato n, Fl., 2003.
2. A. Saltelli. K. Chan, and E. M. Scott (eds.), St'llsitillity Allalysis , Wiley. New York. 2000.
3. v. S. Arpaci, Conduction Heat Trnl1~fer, Addison-Wesley, Reading, MA, 1966.
4. J. V. Beck arxl K. 1. Amo!d, PUn/mner Estimarioll ill El1gil1eeril1g (II1d Science, Wiley. New York, 1977.
5. 1. V. Beck, B. F. Blackwell, and C. R. 51. Clair, f1lver.~e Heal COliduC/ioll, Wiley, New York, 1985.
6. R. L. McMaster. K. 1. Dowding, 1. V. Beck, and D. H. Y. Yen, Methodology to Generate Accurate Solu-
tions for Verification of Three Dimensional Heat Conduction, Mlmer. Heal Trallsfer, Pan 8, 41. 52 1- 541
(2002)
7. J.R. R.A. Martins, A Guide to the Complex-step Derivative Approximation. World Wide Web page,
http://mdolab.utias.utoronto.calcomplex-step.html.
8. 1.R.R.A. Martins, P. Sturdza, and 1.1. Alonso, lbe Complex-step Derivative Approximation, ACM Trml!).
MMh. Soft .....ure. 29. 245-262 (2003).
9. 1. C. Newman and D. L. Whitfield, A Step-size Independent Approach for MultidisCiplinary Sensitivity
Analysis and Design Optimization. AIAA-99-3101, presented at 17th AJAA Applied Aerodynamic Confer-
ence,281une 11uly 1999.
10. L. Massa and J. M. Janus. Aerodynamic Sensitivity Analysis of Unsteady Turbine Stages. A1AA-2001-
2579, presented at 15th AIAA Computatjonal Fluid Dynamics Conference, Anaheim, California, I !~1 4
June 2001.
I J. C. Bischof. P. Khademi, A. Mauer. and A. Carle, Adifof 2.0: Automatic Differentiation of Fortran 77
Programs, IEEE Comput. Sci. Ellg. 18-32 (1996).
12. B. F. Blackwell. K. 1. Dowding. R. 1. Cochran, and D. Dobranich, Utilization of Sensitivity Coefficients
10 Guide the Design o f a Thennal Batlery. Proc. ASME Heat Trrm.l/er Div.. lITO 361-5, 73 - 82 (1998).
1]. K.1. Dowding. B. F. Blackwell. and R. 1. Cochran, Application of Sensitivity Coefficients for Heat Con-
ductio n Problems, Numer. Heat Tranjfer, Pan H, 36, 33-55 (1999).
14. B. F. Blackwell. K. 1. Dowding. and R. 1. Cochran, Development and Implementation of Sensitivity Coef-
ficient Equations for Heat Conduction Problems. Numer. Heat Trallsfer, Pari B. 36. 15 - 32 (1999).
15. H. Hristova. O. Tremblay. S. Etienne. D. Pelletier. and J. Borggaard. Application of a Sensitivity Equation
Method to Transient Non-linear Heal Conduction. AIM 2004-0495. presented at 42nd AfAA Aerospace
Sciences Mccting and Exhibit, January 20)4,
[6. J. N. Mahieu. D. Pelletier. 1. Y. Trepanier. and 1. Borggaard, Second Order Sensitivity and Uncertainty
Analysis for Laminar Airfoil Design. AIAA 2004-742. presented at 42nd AIAA Aerospace Sciences Meeting
and Exhibit January 2004.
17. M, Kleiber and T. D. Hien, The Stochaslic Fillile Elemellt Method, Wiley. New Yurk, 1992.
18. M. Kleiber, H. Autunez, T. D. Hien. and P. Kowalczy k. Parameter Sellsitivities in NOlilillear Mechanics.
Theory and Fillite Elemelll COmpUWtlOlI.l", Wiley. New York, 1997.
19. U. Kirsc h. SITllcluml OplimiZfllion. Springer-Verbg, Berlin. 1993.
REFERENCES 469

20. H. W. Coleman and W. G. Steele, Experimentation alld Vllcertaimy Allalysis for Engineers. 2nd ed ..
Wiley. New York. 1999.
21. 1'. D. Fadale. VllcerwilllY Analysis Usillg Finite Elements. Thesis. University of Washington, Seattle. 1993.
22. T. D. Fadale and A. F. Emery, Effects of Uncertainties on the Sensitivities of Temperature and Heat Fluxes
Usi ng Stochastic Finite Elements, ASME J. Heal Trw!.ifer, 116, 808 -8 14 (1994).
23. 1. C. Helton and F. J. Davis, Sampling-based Melhods for Uncertainty and Sensitivity Analysis. Tech. Rep.
SAND99·2240. Sandia National Laboratories, July 2000.
24. M. D. McKay. W. 1. Conover, and R. 1. Beckman. A Comparison of Three Methods for Selecting Values
of Input Variables in the Analysis of Output from a Computer Code, Tec/momelrics. 21. 239-245 (1979).
25. J. C. Helton and F. J. Davis. Latin Hypercube Sampling and tbe Propagation of Uncertainty in Analyses
of Complex Systems, ReliabililY Eligilleerillg ,md Syslem Safery, 81. 23-69 (2003).
26. R L. Iman and M. J. Sbortencaricr. A FORTRAN 77 Program and User's Guide for the Genemtion of
Latin Hypercube and Random Samples for Use with Computer Models. Tech. Rep. SAND83·2356. Sandia
National Labomtories. March 1984.
27. G. D. Wyss and K. H. Jorgensen. A User's Guide 10 LHS : Sandia 's Latin Hypercube Sampling Software ,
Tech Rep. SAND98·021O, Sandia National Laboratories. February 1998.
28. M. S, Eldred. A A . Guinta, 8. G. van Bloemen Waandcrs, S. F. Wojtkiewicz, W. E. Hart, and M. P.
Alleva. DAKOTA. A Multilevel Pumllc1 Objcct-oriented Fmmework: for Design Optimization. Parameter
Estimation, Uncertainty Quantification. and Sensitivity Analysis. Version 3,1 Users Manuul, Tech. rep.
SAND2001-3796, Sandia N1Hional Laooratories, April 2003.
29, G. Strang, IntroduClion 10 Lineur Algebra, 2nd ed .. Wellesley-Cambridge Press. Wellesley, MA. 1998.
30. R. B. Dowdell. ASME J. F/uid.f Ellg. 104 ,258 (1982).