Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

www.elsevier.com/locate/cam

Operator splitting approach applied to oscillatory flow and heat

transfer in a tube
R. Widuraa , M. Lehna , K. Muralidharb , R. Scherera,∗
a Institut für Praktische Mathematik, Universität Karlsruhe (TH), Germany
b Department of Mechanical Engineering, Indian Institute of Technology Kanpur, India

Received 26 April 2006; received in revised form 7 November 2006

Abstract
The method of operator splitting is applied to an advection–diffusion model as it occurs in a pulse tube. Firstly, the governing
equations of the simplified model are studied and the mathematical description is derived. Then the splitting approach is used to
separate the advection and the diffusion part. Now it turns out that the advective part can be solved analytically and therefore
the computational cost are reduced and the accuracy is increased. It is shown that the method can model an effect called Taylor
dispersion. Applying a domain decomposition strategy, the solution process can be decoupled, reducing the numerical cost even
more. This procedure allows to study the relevant parameters within the model with the goal to maximize the amount of energy
stored within the tube wall. As a measure of efficiency, the amount of energy transferred between the fluid phase and the wall is
chosen.
© 2006 Elsevier B.V. All rights reserved.

MSC: 65M06; 35L65; 80A20

Keywords: Operator splitting; Domain decomposition; Taylor dispersion; Recuperator; Pulse tube

1. Introduction

The physical system under consideration of the mathematical modelling in this work is the so-called pulse tube as
it occurs in Stirling cryocoolers as described by [11]. In this tube, hot and cold fluid oscillates in order to generate
very low temperatures by exploiting several thermodynamical effects in combination with the conservation laws. The
main components of a pulse tube cryocooler (PTC) are a compressor, a regenerator and the pulse tube. The compressor
drives the fluid or gas through the regenerator and the pulse tube back and forth. In the regenerator, the fluid exchanges
heat with the material (a porous medium). Inside the pulse tube, the compression and expansion of the fluid takes place
in order to cool down the sample which can be placed at the cold end of the tube. Mathematical descriptions of this
system have been developed by several authors as [12,3]. The model focusses on the pulse tube only, so we can analyze
its ability to act as a regenerator itself. This kind of cryocooler is referred to as recuperator [1]. The advantage of this

∗ Corresponding author.
E-mail addresses: [email protected] (R. Widura), [email protected] (M. Lehn), [email protected] (K. Muralidhar),
[email protected] (R. Scherer).

0377-0427/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2006.11.019
116 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

R+l
Ts
R

r Tf
u

0 L
z

Fig. 1. Layer model of the tube.

setting is that the number of moving parts in the system is reduced and sedimentation of the porous media, which can
cause severe disturbances and loss of efficiency in todays cryocoolers, is impossible.
The class of advection–diffusion equations can be described in a very general form as introduced in the work by [8]:

ut + ∇ · f(u) = u, u(x, 0) = u0 (x)


with x ∈ Rm , f(u) = (f1 (u), . . . , fm (u)), ∇ = (j/jx1 , . . . , j/jxm ) and  = mi=1 j /jxi . The function f determines
2 2

the advective and the parameter  the diffusive influence on the system. Due to the importance of these equations, many
approaches have been developed to solve them efficiently and effectively. A very common way to address the problem
numerically is to use a finite difference method, which approximates the derivatives by a difference term with finite
stepsize (see [6]). Depending on the stepsize, error and stability properties can be studied.
In this work the method of operator splitting is applied to the advection–diffusion equation [9]. The basic idea is to
split a given differential equation

ut = (A + B)u, u(x, 0) = u0 (x)

with different operators A and B into separate parts for one small splitting step , i.e.,

u∗t = Bu∗ , u∗ (x, 0) = u0 (x),

u∗∗ ∗∗
t = Au , u∗∗ (x, 0) = u∗ (x, ).

The initial value of the second equation is the solution at the time step  of the first line. The step size  has to be chosen
reasonably small in order to guarantee the stability of the whole scheme [4]. The benefit of this method is the reduction
of complexity in the equation and the possibility to apply different schemes to each of the separated parts. A negative
aspect is that a -dependent splitting error may occur in the case of non-commutative operators. A general overview on
operator splitting algorithms for advection dominated advection–diffusion equations is given by [5].
In the following sections we will present the simplified model of a pulse tube and state the governing equations as
well as the initial and boundary conditions. Then we will discretize the system and apply the operator splitting method
in order to reduce the complexity of the problem. While setting up the model for the PTC, we make the following
assumptions as suggested by [3]:

• The flow is laminar, incompressible and has constant thermophysical properties.

• There are no phase differences in the oscillations of the flow and no work is caused by the expansion/compression.
• The convective heat resistance between wall and fluid is far larger than the internal conductive resistance of the tube
wall.
• The heat capacity of the tube wall is much larger than that of the fluid.
• The flow velocity along the radial direction is zero.

We will exclusively analyze the pulse tube itself, neglecting any effects from the piston or other sources. As shown in
Fig. 1, the tube is divided into fluid and wall part. The temperature distribution is assumed to be symmetric with respect
to the angular direction, thus it suffices to consider only the radial cross-section of the tube. Due to the cylindrical shape
of the tube, it is quite natural to introduce cylindrical coordinates to describe the spatial quantities.
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 117

Table 1
List of dimensional and non-dimensionalized quantities

Quantity Unit Description

f 1/s Oscillation frequency

f kg/m3 Fluid density
s kg/m3 Wall density
u m/s Fluid velocity
 Pa s Viscosity
kf W/(m K) Fluid thermal conductivity
ks W/(m K) Wall thermal conductivity
cP J/(kg K) Fluid specific heat capacity
cs J/(kg K) Wall specific heat capacity
L m Tube length
l m Wall thickness
r m Tube diameter
Ta K Ambient temperature
TH K Hot end temperature
TC K Cold end temperature
T −TC
T̂ := TH −TC – Temperature
L
L̂ := R – Length of the tube
ẑ := Rz – z-length scale
 := Rl – Tube thickness
r̂ := Rr – r-length scale

tˆ := Rf2 t – Time
ˆ := R 
2
 f
– Oscillation frequency
u
û := umax – Flow velocity
ks
:= kf – Ratio of wall/fluid thermal conductivity
:= fs – Ratio of wall/fluid diffusivity
P e := R·ufmax – Peclet number

2. The mathematical model

After introducing the non-dimensional variables in Table 1, we will now state the governing equations in non-
dimensional form. For brevity, a slightly different notation than in Table 1 is used by dropping the ‘hats’ (ˆ) and using
the non-dimensional quantities as long as it is not stated otherwise. It is assumed that the only effect within the wall is
the diffusion. Due to the geometrical setting, cylindrical coordinates are used to describe the equations from now on.
Therefore we write
 
jTs 1 jTs j 2 Ts j 2 Ts
= + + (1)
jt r jr jr 2 jz2
for the governing equation of the solid phase. Within the fluid, there exists an advective and a diffusive part. The
advective term describes the inflow of hot or cold fluid and the diffusive term the diffusion that takes place within the
fluid phase
 
jTf jTf 1 jTf j 2 Tf j 2 Tf
+ Pe · u = + + . (2)
jt jz r jr jr 2 jz2
Note that for Peclet numbers P e>1 the diffusion dominates, for P e?1 the advective part is dominant. In PTCs, we
usually deal with advection-driven systems, therefore the main focus lies on high Peclet numbers in the numerical
experiments. Now we will derive the appropriate initial and boundary conditions. The numbers in the following
explanations refer to Table 2. The tube itself is assumed to be adiabatic, which means that there are no fluxes at the
118 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

Table 2
Non-dimensionalized initial and boundary conditions

Wall I.C. Ts (z, r, 0) = 0 

jTs 
B.C. (i) Ts (z, 1, t) = Tf (z, 1, t) (ii) jr r=1+ = 0
jT s jTs
(iii) jz (0, r, t) = 0 (iv) jz (L, r, t) = 0

Fluid I.C. Tf (z, r, 0) = 0  

k f jT f  jT  jTf 
B.C. (i)  = jrs  (ii)  =0
ks jr r=1 r=1 jjrTf r=0
Tf (0, r, t) = 1, u  0 jz (L, r, t) = 0, u0
(iii) jTf (iv)
jz (0, r, t) = 0, u < 0 Tf (L, r, t) = 0, u<0

Fig. 2. Non-dimensionalized boundary conditions for wall and fluid.

outside of the wall (ii). Furthermore, the ‘no-flux’ condition is applied at z = 0 and L (iii), (iv). For the interface
between wall and fluid, it is quite reasonable to assume that the temperature should be the same, so that there is no
discontinuity in the temperature distribution of the whole system (i). For the fluid phase, an interface condition should
also be derived. With respect to the different thermal properties of the fluid and the wall, a restriction to the flux between
the two media is given (i). At the center of the tube, which is due to the symmetry property at r = 0, the flux is again
zero (ii). As outlined in the section on the PTC, the flow through the tube is of oscillatory nature. Therefore it has
to be distinguished between a positive (that is set by convention to be from left to right) and an (opposite) negative
flow. The flow is constant and radial components are neglected. Now, if u is positive, the flow entering at z = 0 is,
again by convention, assumed to have the temperature TH . For negative u the temperature cannot be predicted at that
point and therefore the no-flux condition is set (iii). At z = L, the corresponding statement holds with temperature TC ,
respectively (iv).
In Fig. 2 the non-dimensionalized quantities are introduced and the boundary conditions are included into the overall
model of the tube. In combination with the governing equations, the problem now is well posed as there are four
boundary conditions for the wall as well as for the fluid, corresponding to each first and second order spatial derivative
in r- and z-direction.

3. Discretization methods

Time discretization: For the time-marching the forward time scheme

n+1 n
jT Ti,j − Ti,j
= + O(t) (t → 0)
jt t
is used. Although only of consistency order one, it proves to be robust and stable. Furthermore, the time steps have to
be chosen quite small to satisfy the CFL condition [9]. This leads to a reasonable accuracy even for schemes with a
consistency order of one.
Discretization of the advection term: One possibility to discretize the advective term is the forward space method
of order one. To obtain a better accuracy without using a finer grid, an implicit Crank–Nicholson scheme (see [6]) is
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 119

used in the numerical experiments. This discretization is only performed to simulate a conventional method in order to
compare the results of the operator splitting method with a common conventional approach. The discrete description is
⎧ n+1 n+1 n −Tn
⎪ ui,j Ti,j − Ti−1,j Ti,j i−1,j
⎪
⎪ + + O((z)2 ), ui,j 0
jTf ⎨ 2 z z
u = n+1 n+1 n n (z → 0). (3)
jz ⎪ ui,j Ti+1,j
⎪ − Ti,j Ti+1,jj − Ti,j
⎪
⎩ + + O((z) ), ui,j < 0
2
2 z z
Discretization of the diffusive terms: For the same reasons as in the advection term, we will use the implicit
Crank–Nicholson scheme for the diffusion as well. To discretize the second order derivatives, the centered space
approach
n+1 n+1 n+1 n n +Tn

j2 T 1 Ti+1,j − 2Ti,j + Ti−1,j Ti+1,j − 2Ti,j i−1,j
= + + O((z)2 ) (z → 0), (4)
jz2 2 (z)2 (z)2
n+1 n+1 n+1 n n n

j2 T 1 Ti,j +1 − 2Ti,j + Ti,j −1 Ti,j +1 − 2Ti,j + Ti,j −1
= + + O((r)2 ) (r → 0) (5)
jr 2 2 (r)2 (r)2

is used. The first derivative in radial direction is somewhat special, because the term 1/r arises. Nevertheless, the
centered space discretization is also applicable in this case and we get
n+1 n+1 n n

1 jT 1 Ti,j +1 − Ti,j −1 Ti,j +1 − Ti,j −1
= + + O((r)2 ) (r → 0). (6)
r jr 2rj 2r 2r

While trying to apply these discretization methods to the governing equations for wall and fluid as stated in (1) and
(2), respectively, the resulting linear system will become very large and thus difficult to be solved using a numerical
solver. Therefore, the following way is suggested: apply the operator splitting first, and then the domain decomposition
to reduce the order of the system.

4. Operator splitting

As described by [10] we can split up the advection–diffusion equation of the fluid phase (2) into the advective and
diffusive part
Advection: Tt + uP e∇T = 0, (7)
Diffusion: Tt = T . (8)

Now the decisive assumption under physical considerations is made. We assume that there is no radial advection, the
advection step therefore reduces to
jT
Tt + uP e = 0, (9)
jz
where the velocity u is assumed to be constant. This is reasonable, because the dominant fraction of the advection is the
axial one, due to the high velocities occurring in the pulse tube. If we further assume an initial condition T0 (z)=T (z, 0),
then the analytical solution of this one-dimensional advection equation is known to be

T (z, t) = T0 (z − uPe t). (10)

The application of the operator splitting method is not just intended to reduce the computational cost, but in this case
also enables us to solve one part of the original problem analytically. For the solution of the diffusion problem, we use
some discretization methods described earlier, because the analytical solution is not known.
With this approach we can now assume that the solution of the advection step is free of any errors (besides rounding
errors which can be neglected). As described by [15], we can then find an estimate for the error caused by solving
120 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

0.3 0.3
Pe=1
Pe=10
0.2 Pe=100 0.2
Pe=1000

0.1 0.1
Error

Error
0 0

-0.1 -0.1

-0.2 -0.2

-0.3 -0.3
0 1 2 3 4 5 0 1 2 3 4 5
z z

Fig. 3. Error for P e = 1, 10, 100, 1000 at t = 0.004, t = 0.0001, z = 0.01. (a) Error between analytical and conventional method. (b) Error
between analytical and operator splitting method.

∂Ω2

1+δ

Ω2
1
Ω1 Γ
r

0 L
z ∂Ω1

Fig. 4. Domain decomposition into fluid domain 1 and wall 2.

the diffusion numerically. To measure this error, we use the local truncation error as described by [9] and estimate the
maximal error in z-direction
     
1 1  1 3 
LD
(., t)     − C  · t +   C  · (t)2 (11)
t ∞ 2 6p  6 

with the diffusion coefficient  ( = 1 in this case), p = (t)/(x)2 = const and C some constant. The reduction of
the error can be seen in Fig. 3, which shows the absolute error between the analytical solution of the one-dimensional
advection–diffusion equation and the conventional method on the left side and the operator splitting method on the
right side, respectively. This error description corresponds to the solutions plotted in Fig. 5. As mentioned earlier, one
problem while using operator splitting is the splitting error. But in the case of the advection–diffusion equations as
stated here, this error vanishes due to the commutativity of the operators as shown by [9].
Domain decomposition: In order to simplify computations and to decouple the domains of wall and fluid, there exist
several schemes, e.g., the implicit generalized Schwarz alternating method [13] or the Uzawa algorithm (see [14,2]).
Both methods are iterative and are repeated until sufficient convergence occurs or other conditions force the procedure
to stop. The main advantage of domain decomposition methods is on the one hand, that it reduces the order of the linear
systems that have to be solved and therefore reduces computational cost. On the other hand, the separation of the two
parts gives the opportunity to distribute the computations on several processors. Thus, it is very desirable to decompose
the system in order to exploit these advantages. In Fig. 4 we introduce a very natural way of domain decomposition
into fluid 1 and wall 2 . Note that 1 ∪ 2 = . Furthermore, we let j 1 and j 2 be the corresponding boundaries of
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 121

Table 3
Complete algorithm in pseudocode

1. Split equation into advection and diffusion

2. Perform for every time step t:
(a) Solve diffusion:  
jT1  jT2 
jr  = − jr  = 
Initial guess 0

Repeat {
solve MT n+1
1 = f1 on 1
jT1n+1
with jr
= i on

solve MT n+1
2 = f2 on 2
jT2n+1
with jr
= −i on

S1i and S2i being the corresponding solutions, set

y1i := S1i |
y2i := S2i |

then update
i := i−1 + (y2i − y1i ) with 0 <  < 1
}
Until |i − i−1 | 
(b) Apply advection using analytical solution and save current temperature profile
3. Plot results

the domains. Then = j 1 ∩ j 2 is their interface. We also introduce the vectors n1 and n2 which are outside normal
to 1 and 2 , respectively. When applying the domain decomposition method, the main challenge is the treatment
of this interface. Several possibilities are discussed by [7]. In our computational experiments, we will use the Uzawa
algorithm. In order to solve the overall problem

MTn+1 = f

with the corresponding initial and boundary conditions. The whole algorithm, containing the Uzawa domain decom-
position method, is shown in Table 3. The algorithm starts with splitting the original equation into an advective and
a diffusive part. To solve the diffusion equation for fluid and wall, the Uzawa algorithm is applied. The value for the
interface flux jT1 /jr| = −jT2 /jr| = 0 has to be guessed first. Then the problem is solved for the two subdomains
1 and 2 with the corresponding boundary conditions on j 1 ∩ j , j 2 ∩ j , respectively, and . The solution of
both problems at the interface is then used to update the interface flux i . This process is repeated until the change of
i is sufficiently small (i.e., y1i and y2i are similar). The analytical solution of the advection equation is known, so that
it can be applied directly to the resulting solution of the diffusion. The temperature data for the entire tube is saved,
and then, the next time step is performed and the iteration starts again. The last step is to plot the results.

5. Numerical results

The results of the operator splitting method are compared to a conventional as well as the analytical solution, showing
that the operator splitting algorithm gives more accurate and stable results. In the two-dimensional case, the operator
splitting is even more superior to the conventional method which struggles with heavy oscillations. The behaviour of
the flow in the full model is analyzed. There we see the different temperature distributions, depending on time and
position in space. Then, the amount of energy stored in the wall is computed for different parameters and the optimal
ratio of stored and released heat is found for high Peclet numbers and low frequencies.
122 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

1.2 1.2
Conv.
OS
1 Analytical 1

0.8 0.8

0.6 0.6

T
T

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
z z

1.2 1.2

1 1

0.8 0.8

0.6 0.6
T

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
z z

Fig. 5. Temperature profile at t = 0.004, t = 0.0001, z = 0.01. (a) Pe = 1, (b) Pe = 10, (c) Pe = 100 and (d) Pe = 1000.

One-dimensional case: As a reference problem, the one-dimensional unsteady advection–diffusion equation with
constant velocity u = 1
jT jT j2 T
+ Pe = 2 (12)
jt jz jz
and initial condition T (x, 0)=0 and the two boundary conditions T (0, t)=1 as well as limz→∞ T (z, t)=0 is considered.
The analytical solution of Eq. (12) is given by
 ∞  
2 Pe z2
T (z, t) = √ exp − + 2
(z − 2 ) d. (13)
 z/(2√t) 2 8
Fig. 5 shows that the operator splitting and the conventional method, both approximate the true solution very accurate
for low Peclet numbers. In the case of high Peclet numbers, the operator splitting approach clearly outperforms
the conventional method. The grid used in this case was t = 0.0001 (time discretization) and z = 0.01 (spatial
discretization). In pulse tube coolers high Peclet numbers are quite common, which implies the advantage of using an
operator splitting method compared to other schemes.
Two-dimensional case: To verify the operator splitting method in the two-dimensional case, the solutions of the
conventional approach are compared to the operator splitting algorithm. Fig. 6 shows that the conventional method
generates heavy oscillations, whereas the operator splitting method shows no disturbances at all. The length of the tube
is L = 5 and the grid is chosen to be t = 0.01, r = z = 0.1. The splitting method gives accurate results from the very
beginning with no oscillating errors, which can be observed when using conventional methods. Therefore the operator
splitting algorithm is superior to conventional methods in this particular case.
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 123

1.4 1.4
Upwind
OS
1.2 1.2

1 1

0.8 0.8
T

T
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t t

1.4 1.4

1.2 1.2

1 1

0.8 0.8
T
T

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t t

Fig. 6. Bulk mean temperature at z = 1. (a) Pe = 1, (b) Pe = 10, (c) Pe = 100 and (d) Pe = 1000.

6. Taylor dispersion

After validating the code, the effect of Taylor dispersion is studied. When a fluid flows through a tube, a velocity
profile develops. This profile is not uniform over the cross-section of the tube, at some radial positions the fluid flows
faster than at others. The laminar profile of Newtonian fluids has a parabolic velocity distribution. Apart from the
purely convective dispersion, diffusion can also be of influence. In our case, the hot inflow hits the cold fluid inside
the tube which causes the temperature to disperse, but it also induces radial composition gradients. This leads to
diffusive fluxes at the frontside of the inflow. This flux becomes important, if the radial diffusive fluxes are roughly
of the same order of magnitude as the convective axial fluxes. This is the case, when either the axial velocity is
very low or when the radial distances are very small. The operator splitting method is capable of making the Taylor
dispersion visible. The temperature development at one specific point in the tube (in this case z = 1) is tracked over
time. These curves for a theoretical flat velocity profile are compared with the real parabolic profile. The latter profile
is defined as u(r) = 2umax (1 − r 2 ). Therefore, the fluid at r = 0 travels much faster compared to the constant velocity
u = 1 of the ‘flat’ profile, but close to r = 1, the parabolic velocity is slower. Thus the wave of the incoming fluid
hits some areas earlier, others later. It is expected that the temperature increase at the fixed point generated by the
parabolic profile will be ‘broader’ than the one generated by the flat velocity profile. As we can see in Fig. 7, the
expectations are verified by the numerical results computed for different Peclet numbers. Especially, in the case of
high Peclet numbers there is a distinctive effect of the Taylor dispersion. This effect is similar to a change of the
Peclet number itself. Therefore the velocity u can be kept constant, while modifying the Peclet number accordingly.
That means a parameter (Pe) is searched such that the problem (while introducing the explicit r dependence of the
124 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

1 1

0.8 0.8

0.6 0.6
T

T
0.4 0.4

0.2 0.2
Flat profile
Parab. profile
0 0
0 1 2 3 4 5 0 0.1 0.2 0.3 0.4 0.5
t t

1 1

0.8 0.8

0.6 0.6
T

0.4 0.4

0.2 0.2

0 0
0 0.01 0.02 0.03 0.04 0.05 0 0.001 0.002 0.003 0.004 0.005
t t

Fig. 7. Bulk mean temperature at z = 1 for flat and parabolic velocity profile. (a) Pe = 1, (b) Pe = 10, (c) Pe = 100 and (d) Pe = 1000.

velocity u)
 
jT jT 1 jT j2 T j2 T
+ Pe · u(r) = + 2 + 2 (14)
jt jz r jr jr jz
can be rewritten with constant velocity, e.g., u = 1, and parameter (P e)
 
jT jT 1 jT j2 T j2 T
+ P e · (P e) = + 2 + 2 . (15)
jt jz r jr jr jz
In order to find (Pe), the optimization problem with respect to Pe that best fits the two curves together has to be
solved. This is a very important possibility in real applications, where often the transition parameter (P e) can only be
guessed upon experimental data. The method presented gives the opportunity to calculate that parameter analytically,
only dependent on the discretization and splitting errors.

7. The complete setting

The previous parts present that the operator splitting produces accurate results and that our model describes the
physical processes as expected. We will therefore now analyze the complete setting including the following aspects:

• Oscillatory flow.
• Advection and diffusion within the fluid.
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 125

Pe=1
Pe=10
Pe=100
Pe=1000

Fig. 8. Plot key.

a b
4 4
3 3
2 2
1 1
λ∗(dT/dr)

λ∗(dT/dr)
0 0
-1 -1
-2 -2
-3 -3
-4 -4
6.9 6.95 7 6.9 6.95 7
t t

c d
4 4
3 3
2 2
1 1
λ∗(dT/dr)

λ∗(dT/dr)

0 0
-1 -1
-2 -2
-3 -3
-4 -4
6.9 6.95 7 6.9 6.95 7
t t

e f
4 4
3 3
2 2
1 1
λ∗(dT/dr)

λ∗(dT/dr)

0 0
-1 -1
-2 -2
-3 -3
-4 -4
6.9 6.95 7 6.9 6.95 7
t t

Fig. 9. Interface flux for oscillatory flow (1/ = 0.014). (a) Hot cycle,  = 0.65, (b) cold cycle,  = 0.05, (c) hot cycle,  = 0.1, (d) cold cycle,
 = 0.1, (e) hot cycle,  = 0.15 and (f) cold cycle,  = 0.15.
126 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

• Diffusion within the wall.

• Heat flux at the interface between fluid and wall.

In addition to the issues already discussed in this section, a solution method for the wall temperature as well as for the
interface has to be implemented. The principle of a pulse tube cooler is based on its ability to store energy (heat) at
one time and give it back to the fluid at another time. In todays cryocoolers, this is usually done using a regenerator
mesh. Due to the problems that may arise using such a regenerator, we want to find out if the wall itself can store
enough energy in order to act as a regenerator. This kind of PTC is generally called recuperator. With the developed
model and the presented numerical methods, we are able to compute the energy by exploiting that the temperature
within the wall is (approximately) known. To study this issue we will compare the averaged interface temperature and
flux between fluid and wall. By independently varying the Peclet number Pe, the wall thickness  and the frequency
 (which corresponds to the cycle length 1/) we find the optimal parameters for which the amount of energy stored
within the wall is maximized. The regular operation of the device is usually of interest and thus, the data at some
time instant at which the cycles are balanced and in a quasi-steady state is used. The time when this state is reached
depends on the wall thickness, the frequency and the Peclet number. The flux dT /dr always decreases until it reaches
the balanced state where the amplitude is constant and the flux varies between fixed values. This behaviour is to be
expected because at the beginning the wall has the temperature TC = 0. By letting hot fluid rush into the tube, the
flux at the interface is very high until the tube wall gradually heats up. In the quasi-steady state, the heating effects
by the hot and the cooling effects by the cold fluid are equal, and therefore the mean flux does not change anymore.
We can furthermore observe that the quasi-steady state is reached faster for high Peclet numbers. This is reasonable
as well, because high Peclet numbers correspond to high velocities, and thus the whole tube is flooded with hot
fluid leading to a faster exchange of heat, i.e., heating up the wall and therefore reaching the balanced state. Under
ideal conditions, the temperature profile along the tubes length within the fluid would be linear, falling from unity to
zero. The extent of non-ideal conditions that arise when the tube wall is used to store energy can be seen in Fig. 10.

1 1

0.8 0.8

0.6 0.6
T

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
z z

1 1

0.8 0.8

0.6 0.6
T

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
z z

Fig. 10. Axial temperature profiles for oscillatory flow for Pe = 100.
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 127

4 4
3 3
2 2
1 1
λ∗(dT/dr)

λ∗(dT/dr)
0 0
-1 -1
-2 -2
-3 -3
-4 -4
6.5 6.6 6.7 6.8 6.9 7 6.5 6.6 6.7 6.8 6.9 7
t t

5 4
4 3
3 2
2 1
λ∗(dT/dr)

λ∗(dT/dr)

1 0
0 -1
-1 -2
-2 -3
-3 -4
-4 -5
6.5 6.6 6.7 6.8 6.9 7 6.5 6.6 6.7 6.8 6.9 7
t t

Fig. 11. Interface flux for oscillatory flow ( = 0.1). (a) Hot cycle, 1/ = 0.014, (b) cold cycle, 1/ = 0.014, (c) hot cycle, 1/ = 0.05 and (d) cold
cycle, 1/ = 0.05.

The figure shows the temperature profile at four different time instances for a Peclet number of Pe = 100 and with a
wall thickness of  = 0.01. The first two pictures were captured during the hot cycle, the other two during the cold
cycle. For this calculations as well as for all the following ones, the grid t = 0.002, z = 0.2, r = 0.1, rw = 0.002
was used, where rw corresponds to the grid size used within the tube wall. For the Uzawa domain decomposition
algorithm, the convergence parameters  = 0.2, tol = 0.00001 and a maximal iteration number of 1000 were used.
The starting effects of the cryocooler can be explained like above. A real device will work most of the time in the
balanced state. Therefore we will concentrate on effects arising in this quasi-steady case. Thus, we often leave out the
first time units and directly look at the behaviour at a later stage, usually between 6 and 7 time units. For brevity we also
drop the key in the following plots and always use the labels as shown in Fig. 8. The mean flux at the interface is shown
in Fig. 9 for 1/ = 0.014. Here the effects at the interface are hardly influenced by the wall thickness. Surprisingly,
the flux for thinner tubes is even a little bit higher than it is for thicker tubes. This very small effect might be caused by
numerical errors. Besides the positive (negative) flux in the hot (cold) cycle we always find a small flux at the beginning
of each new cycle, which is opposite to the main flux. The ‘new’ flow rushing in has to substitute the fluid that filled
the tube up to that moment. This takes some time depending on the Peclet number and until this process is finished, the
‘old’ fluid contributes to the averaged interface flux and causes it to change its sign. We note that this effect caused by
the opposite flux is also not much affected by the wall thickness. A much stronger effect is observed when changing the
frequency . The fluid in the corresponding cycle has more time to transfer energy to or from the wall and therefore
changes the temperature at the interface. Thus, the flux also varies strongly with the frequency. As we see in Figs. 11
and 12 not only the duration of heat transfer increases with decreasing frequency, but also the absolute value of the flux
for  = 0.1. Due to the stronger change in temperature at the interface and the increased time to spread this temperature
128 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

5 4
4 3
3 2
2 1
λ∗(dT/dr)

λ∗(dT/dr)
1 0
0 -1
-1 -2
-2 -3
-3 -4
-4 -5
6.5 6.6 6.7 6.8 6.9 7 6.5 6.6 6.7 6.8 6.9 7
t t

7 2
6 1
5 0
4 λ∗(dT/dr) -1
λ∗(dT/dr)

3 -2
2 -3
1 -4
0 -5
-1 -6
-2 -7
5.5 6 6.5 7 5.5 6 6.5 7
t t

Fig. 12. Interface flux for oscillatory flow ( = 0.1) (a) Hot cycle, 1/ = 0.1, (b) cold cycle, 1/ = 0.1, (c) hot cycle, 1/ = 0.5 and (d) cold cycle,
1/ = 0.5.

into the wall, the flux at the beginning of a new cycle is much higher. Besides this, we also observe that the peak of the
opposite flow in Fig. 9 is much smaller compared to the peak of the main flux. The total energy stored and released
from the wall is therefore higher (Figs. 10–12).

8. Conclusion

The previously discussed results are summarized in Figs. 13 and 14. In both cases we integrate the interface flux
over one cycle (hot and cold, respectively) in steady state for various values of  and . The result of this integration
is the amount of energy stored (released) within (from) the wall. In Fig. 13, the flux in the case of  = 0 (that is no
wall at all), has to be 0. Then, with increasing wall thickness, also the energy has to increase but from a certain point
on, the amount of stored energy within a cycle stays almost constant. We remember that the integration is performed
for all fluxes within one cycle, including the one in the opposite direction. The slight decrease of energy after a wall
thickness of  = 0.05 can therefore be explained with a bigger influence of these opposite fluxes compared to the main
flux. In Fig. 14, we see a quite different picture. In the case of infinite frequency  and thus a cycle length 1/ of
zero, the amount of energy exchange in the quasi-steady case becomes zero. This is because the flow inside the tube
oscillates that fast that the positive and the negative flux cancel out each other, neither the hot nor the cold fluid can
really enter and flow through the tube. But with decreasing frequency and again integrating over a whole cycle, we see
an almost linear increase of the stored energy. This leads to the conclusion that, for optimal and most effective energy
storage, the wall thickness should be bigger than 0.02 times the tube radius and the frequency should be rather low to
enable the fluid in the hot as well as the cold cycle to exchange heat with the wall and thus make it an energy storage.
 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130 129

0.035 0.005

0.03 0
-0.005
0.025
-0.01
Energy

0.02

Energy
-0.015
0.015
-0.02
0.01
-0.025
0.005 -0.03
0 -0.035
0 0.05 0.1 0.15 0 0.05 0.1 0.15
δ δ

Fig. 13. Dependence between wall thickness and interface flux. (a) Hot cycle and (b) cold cycle.

a b
1.6 0.2
1.4 0

1.2 -0.2
-0.4
1
Energy

Energy

-0.6
0.8
-0.8
0.6
-1
0.4 -1.2
0.2 -1.4
0 -1.6
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
1/ω 1/ω

Fig. 14. Dependence between cycle length and interface flux. (a) Hot cycle and (b) cold cycle.

Of course, the model used in these calculations was simplified at several points and a deeper physical analysis as well
as systematic experiments have to be performed to elaborate the extent to which a usage is reasonable. But the results
generated with the simulation of an oscillatory flow in a tube show that the wall can be used as an energy storage and
that the benefits might outperform classical cooling devices in specific fields of application.
These results could be achieved due to the exploitation of the operator splitting method, which reduced the compu-
tational cost significantly and thus made the detailed analysis possible.

Acknowledgements

This work is a joined project by the University of Karlsruhe and the Indian Institute of Technology Kanpur. Many
of the results presented here were developed during a research visit of the first author at the Department of Mechanical
Engineering. The stay was generously supported by the Deutscher Akademischer Austausch Dienst (DAAD) and the
Indian Institute of Technology itself. The first author would like to thank Prof. Muralidhar for the invitation and
inspiration. We thank the referees for valuable comments.
130 R. Widura et al. / Journal of Computational and Applied Mathematics 211 (2008) 115 – 130

References

[1] J.G. Brisson, G.W. Swift, Measurements and modelling of recuperator for superfluid stirling refrigerator, Cryogenics 34 (12) (1994) 971–982.
[2] J. Cea, Lectures on Optimization Theory and Algorithms, Springer, Berlin, 1978.
[3] K.S. Cho, D.-Y. Lee, S.T. Ro, Estimation of Stirling engine regenerator on the analogy of laminar oscillating flow in a circular pipe, in:
Proceedings of the 30th Intersociety Energy Conversion Engineering Conference, vol. 3, 1995, pp. 371–376.
[4] P. Deuflhard, A. Hohmann, Numerical Analysis in Modern Scientific Computing, second ed., Springer, New York, 2003.
[5] M.S. Espedal, K.H. Karlsen, Numerical Solution of Reservoir Flow Models Based on Large Time Step Operator Splitting Algorithms, Lecture
Notes in Mathematics, vol. 1734, 1998, pp. 9–77.
[6] J.D. Faires, R.L. Burden, Numerische Methoden, Spektrum Akademischer Verlag, Heidelberg, 1993.
[7] K. Gustafson, Domain decomposition, operator trigonometry, Robin condition, Contemp. Math. 218 (1998) 432–437.
[8] K.H. Karlsen, N.H. Risebro, An operator splitting method for nonlinear convection-diffusion equations, Numer. Math. 77 (1997) 365–382.
[9] R.J. LeVeque, Numerical Methods for Conservation Laws, second ed., Birkenhäuser, Zürich, 1992.
[10] K. Muralidhar, T. Sundararajan, Computational Fluid Flow and Heat Transfer, second ed., Narosa Publishing House, New Delhi, 2003.
[11] A.J. Organ, Thermodynamics and Gas Dynamics of the Stirling Cycle Machine, first ed., Cambridge University Press, New York, 1992.
[12] J.G. Organ, Gas dynamics of the temperature-determined Stirling cycle, J. Mech. Eng. Sci. 23 (4) (1982) 207–216.
[13] T. Sheorey, Modelling of enhanced oil recovery from a porous formation, Ph.D. Thesis, Indian Institute of Technology Kanpur, Department of
Mechanical Engineering, IIT Kanpur, 2001.
[14] H. Uzawa, Iterative Methods for Concave Programming, Stanford University Press, Stanford, CA, 1958.
[15] R. Widura, Applying operator splitting to advection–diffusion equations of oscillatory flows, Master’s Thesis, Universität Karlsruhe (TH),
Fakultät für Mathematik, 2005.