Transport in Porous Media (2021) 137:195–214

[Link]

An Explanation to the Nusselt–Rayleigh Discrepancy

in Naturally Convected Porous Media

Po‑Wei Huang1,2 · Florian Wellmann1,3

Received: 26 April 2020 / Accepted: 26 January 2021 / Published online: 19 February 2021
© The Author(s) 2021

Abstract
We model hydrothermal convection using a partial differential equation formed by Darcy
velocity and temperature—the velocity formulation. Using the Elder problem as a bench-
mark, we found that the velocity formulation is a valid model of hydrothermal convection.
By performing simulations with Rayleigh numbers in the non-oscillatory regime, we show
that multiple quasi-steady-state solutions can be one of the reasons that caused the Nus-
selt–Rayleigh discrepancy found in previous experiments. The results reveal more under-
standings about the nature of uncertainty of convection modes in porous media.

Keywords Natural convection · Entropy production · Basin stability analysis

1 Introduction

Natural convection in porous media occurs in various geological and industrial settings,
such as groundwater, geothermal reservoirs, heat sinks and thermal energy storage. We use
the Rayleigh number (Ra) to characterize the material properties of porous media, such
as permeability, fluid density and the temperature difference between the top and bottom
boundaries. We measure the quality of convective heat transfer using the Nusselt number
(Nu). From an engineering perspective, it is beneficial to correlate the Rayleigh number
and the Nusselt number.
Cheng (1979) compiled the experimental, analytical and numerical results of the
Nusselt number and the Rayleigh number for convection heat transfer in a porous layer
heated from below. The compilation showed widespread Nusselt numbers for a particular

* Po‑Wei Huang
[Link]@[Link]
Florian Wellmann
[Link]@[Link]‑[Link]
1
Computational Geoscience and Reservoir Engineering, RWTH Aachen University, Aachen,
Germany
2
Geothermal Energy and Geofluids Group, Institute of Geophysics, Department of Earth Sciences,
ETH Zürich, Zürich, Switzerland
3
Jülich-Aachen Research Alliance - Center for Simulation and Data Sciences (JARA-CSD),
Aachen, Germany

13
Vol.:(0123456789)
196 P.-W. Huang, F. Wellmann

Rayleigh number. We refer to this phenomenon as the Nu–Ra discrepancy. Experimental

and analytical approaches have addressed this problem. Lister (1990) performed experi-
ments using a large porous slab and showed that lateral thermal dispersion is one of the
reasons that caused the Nu–Ra discrepancy. Vadasz (2010) and Vadasz and Braester
(1992) used analytical techniques to show that boundary and domain imperfections are the
causes of the Nu–Ra discrepancy. Karani and Huber (2017b) conducted pore-scale lattice-
Boltzmann simulations and concluded that thermal disequilibrium between solid and fluid
phases could cause the Nu–Ra discrepancy. Furthermore, Karani et al. (2017) proposed a
fractional-order thermal convection model that captures not only the Nu–Ra discrepancy
but also the advance and delay of the onset of convection. The authors explain this phe-
nomenon by examining different effects that can be introduced by experimental setups.
Therefore, we present numerical simulations, which can be set up as experiments without
the imperfections mentioned above.
In a three-dimensional box of saturated porous media, convection can happen in a two
or three-dimensional setting (Beck 1972). Straus and Schubert (1978) found out that two-
dimensional flows have a larger Nusselt number compared to three-dimensional flows
when Ra ≤ 97 using numerical simulations. The numerical simulations from Holst and
Aziz (1972) and Horne (1979) also produced this effect. Therefore, if both two and three-
dimensional convection can exist in the same box, then the system can at least have two
Nusselt numbers. More promisingly, the simulations of Straus and Schubert (1979) indi-
cate that it is always possible to force either steady two-dimensional or steady three-dimen-
sional convection by proper choice of initial conditions. Govorukhin and Shevchenko
(2017) also showed that the selection scenarios of a convection mode strongly depend on
the initial temperature distribution of the porous media. The cosymmetry theory developed
by Yudovich (1991) showed that there exists an infinite number of stable stationary flow
for a fixed Rayleigh number.
Therefore, we hypothesize that the non-uniqueness of convection modes in a box of sat-
urated porous media is one of the reasons that caused the Nu–Ra discrepancy. To prove this
hypothesis, we perform several numerical simulations using different box sizes and initial
conditions of the temperature field. We use the finite element solver—MOOSE Framework
(Gaston et al. 2009) to simulate the physics of naturally convected porous media.
Acknowledging that a specific natural or artificial porous media can have multiple heat
transfer rates is not too useful for decisions in reservoir engineering. Karani and Huber
(2017a) used the basin stability analysis (Menck et al. 2013) to quantify the probability of
the occurrence of different convection modes in a two-dimensional setting. We employ the
same method for a three-dimensional setting. This probability of convection modes can
provide us more information and obtain the expectation value of the Nusselt number.

2 Materials and Methods

2.1 Governing Equations

We present the conservation laws that model natural convection in porous media under the
following assumptions mentioned by Horne (1979):

• The Boussinesq approximation.

• Inertial effects are small, or low Reynolds number.

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 197

• Viscosity of the fluid is constant.

• Thermal dispersion is negligible.
• Saturating fluid and porous solid are in thermal equilibrium.
∇ ⋅ 𝐪 =0,
√ 𝜕2 T
∇2 u + Ra =0,
𝜕x𝜕y
√ � �
𝜕2 T 𝜕2 T
∇2 v − Ra + 2 =0, (1)
𝜕x2 𝜕z
√ 𝜕2 T
∇2 w + Ra =0,
𝜕y𝜕z
√
Tt + Ra(𝐪 ⋅ ∇T) − ∇2 T =0,

where 𝐪 = (u, v, w), the non-dimensional Darcy fluxes in x, y and z directions. T is non-
dimensional temperature, the subscript t is the partial derivatives with respect to time, and
Ra is the Rayleigh number. Gravity effects act on the vertical direction y. The Rayleigh
number is defined as

𝜌20 g𝛼Kp d𝛥Tcf

Ra = ,
𝜇km

where 𝜌0 is the reference density of the fluid, g is the gravity, 𝛼 is the thermal expansion
rate of fluid, Kp is the permeability, d is the vertical height of the porous media, 𝛥T is the
temperature difference between the top and bottom boundaries, 𝜇 is the dynamic viscosity
and km is the overall thermal conductivity. This formulation of velocity and temperature is
also used by Florio (2014).
The non-dimensional variables are
T =T0 + (𝛥T)T ∗ ,
𝜌f =𝜌0 (1 − 𝛼(𝛥T)T ∗ ),
𝐪 =[q]𝐪∗ , (2)
∗ ∗ ∗
(x, y, z) =d(x , y , z ),
t =[t]t∗ ,
where T0 is the reference temperature at the bottom boundary, 𝜌f is the fluid density. The
non-dimensional variables are defined with asterisks, and we drop them for convenience
throughout this paper. The scalings are chosen such that the square root of the Rayleigh
number exists in both momentum and energy conservation equations
√
[ ] KBgkm
q = ,
𝜇cf d
(3)
𝜙𝜌0 cf d2
[t] = .
km

Since we are using a finite element solver, we require the weak form:

13
198 P.-W. Huang, F. Wellmann

� √ 𝜕𝜔 𝜕T �
∫𝛺
∇𝜔 ⋅ ∇u + Ra d𝛺 = 0,
𝜕y 𝜕x
� √ � ��

∫𝛺
𝜕𝜔 𝜕T 𝜕𝜔 𝜕T
∇𝜔 ⋅ ∇v − Ra + d𝛺 = 0,
𝜕x 𝜕x 𝜕z 𝜕z
� √ 𝜕𝜔 𝜕T � (4)

∫𝛺
∇𝜔 ⋅ ∇w + Ra d𝛺 = 0,
𝜕y 𝜕z
� √ �
∫𝛺
𝜔Tt + 𝜔 Ra𝐪 ⋅ ∇T + ∇𝜔 ⋅ ∇T d𝛺 = 0.

The weak form is derived by multiplying the strong form by a test function 𝜔 and inte-
grate through the simulation domain. We choose a first order Lagrangian basis function for
velocity 𝐪 and second order Lagrangian basis function for temperature T. The time integra-
tion method of the transient solver is the Crank–Nicolson method. We approximate the
CFL number on an element as
|u| + |v| + |w|
Celem = 𝛥t ,
hmin

where hmin is the minimum length of the element w.r.t x, y and z axes. In general, we pick
the maximum CFL number of all elements as the representative CFL number, and we make
sure during each timestep, the representative CFL number is around 0.5. When the entropy
production is constant with respect to time, we claim that the simulation has reached a
steady state (Börsing et al. 2017).

2.2 Benchmark Problems and Results

The benchmark results of our code implementation are presented in this subsection. They
are essential in this work, but do not serve as our main focus.

2.2.1 The Elder Problem

Elder studied convection caused by localized heating with the Hele-Shaw cell experiment
and numerical solutions (Elder 1967). It is one of the benchmark problems in software
such as FEFLOW (Trefry and Muffels 2007; Diersch 2014), SUTRA (Voss 1984) and
HydroGeoSphere (Brunner and Simmons 2012; Simmons and Elder 2017). Therefore, we
benchmark our finite element implementation of the velocity formulation using the Elder
problem.
The velocity formulation in 2D is
√ 𝜕2 T
∇2 u + Ra =0,
𝜕x𝜕y
√ 𝜕2 T
∇2 v − Ra 2 =0,
𝜕x
√
Tt + Ra𝐪 ⋅ ∇T − ∇2 T =0.

The boundary and initial conditions are defined in Fig. 1. Using the physical parameters of
the Elder problem defined by Graf and Boufadel (2011), we have a non-dimensionalized
problem with Ra = 521.3. We use a structured quadrilateral mesh, with the 120 elements

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 199

T = 0, v = 0
∂T
∂x=0 T (x, y, t = 0) = 0 =0
∂T
∂x

1
u=0 u=0
y ∂T
=0 ∂T
=0
∂y T = 1, v = 0 ∂y
x
1 2 1

Fig. 1  The boundary and initial conditions of the non-dimensionalized Elder problem

in the x direction and 32 elements in the y direction. We choose a uniform timestep

𝛥t = 0.0001. We compare our solution with the 2, 5 and 10 years simulation results of
Graf and Boufadel (2011), which corresponds to 0.01, 0.025 and 0.05 non-dimensionalized
time. See Fig. 2 for the benchmark results (Table 1).

2.2.2 Beck’s Box

For a 3D box of saturated porous media of a given dimension, Beck (1972) derived the
preferred cellular mode during the onset of convection. We benchmark our code imple-
mentations using different boxes with lengths h1 and widths h2. We use the notation of [h1,

Fig. 2  Comparison of numerical solutions of the Elder problem using the velocity formulation. Left: refer-
ence solution of Graf and Boufadel (2011). Right: our solution

13
200 P.-W. Huang, F. Wellmann

Table 1  The physical properties Parameter Value

of the porous media. Modified
from Graf and Boufadel (2011).
Porosity (𝜙) 0.1
The thermal expansion rate is
calculated using the density Reference fluid density (𝜌0) 999.526 kg m−3
of water as a function of Reference fluid viscosity (𝜇0) 1.239 × 10−3 kg m−1 s−1
temperature from Thiesen et al. Thermal conductivity of fluid phase (kf ) 0.6 kg m s−3 K−1
(1900), assuming a linear growth
of volume Thermal conductivity of solid phase (ks) 1.588 89 kg m s−3 K−1
Heat capacity of fluid phase (cf ) 4184 m2 s−2 K−1
Heat capacity of solid phase (cs) 0 m2 s−2 K−1
Thermal expansion rate (𝛼) 1.6163 × 10−4

h2] to represent the scaled box dimensions normalized by the vertical height, as illustrated
in Fig. 4. The boxes we choose are [1.2, 1.2], [2.0, 0.5] and [3.0, 1.0]. The Rayleigh num-
ber is set as 42.25, slightly above the critical Rayleigh number, such that the system starts
convecting. The initial condition of temperature is the conductive solution with a ±1% per-
turbation. We use the notation of (m, n) to represent the cellular modes. The results are in
Table 2, and they agree with Beck’s analytical cellular modes.

2.3 Quality Measures of Convective Heat Transfer

The Nusselt number and entropy production are used to measure the quality of convective
heat transfer in our simulations. The Nusselt number is defined as the ratio of total heat
transfer and the stationary conductive heat transfer. Due to simplicity and practical rea-
sons, the Nusselt number is widely used in experiments as a quality measure of convection.
However, in numerical simulations, we have access to values of temperature and Darcy
fluxes in space. Therefore, we can use another quality measurement—entropy production.
Entropy production is generally a better assessment of convection, due to its thermody-
namic considerations (Herwig 2016). The Nusselt number is a combination of the quality,
given by the entropy production, and the quantity of heat transfer, given by the heat flux
involved (Herwig 2016). Though, in order to revisit the Nusselt–Rayleigh relation and gain
new insights from it, our simulations still contain the calculation of the Nusselt number
h2 h1

h1 h2 ∫0 ∫0
1 𝜕T |
Nu = − | dxdz, (5)
𝜕y |y=1

Table 2  Comparison of Beck’s cellular mode and the simulated cellular mode.
Box type [h1 , h2 ] Rac Beck’s cellular mode Simulated cellular mode

[1.2, 1.2] 40.553 (1, 1) (1, 1)

[2.0, 0.5] 39.478 (2, 0) (2, 0)
[3.0, 1.0] 39.478 (0, 1) or (3, 0) (0, 1) or (3, 0)

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 201

where h1 and h2 are the length of the box with respect to x and z-direction. This is analo-
gous to the calculations of Hewitt et al. (2014), except that we utilize heat flux on the top
boundary.
Bejan (2013) formulates the volumetric rate of the total entropy production in a fluid-
saturated porous medium as a result of both irreversible heat transfer (subscript therm) and
fluid flow friction (subscript visc)
��
dStotal
(6)
�� �� ��
= Ṡ total = Ṡ therm + Ṡ visc ,
dt
with Ṡ ( J s−1 m−3 K−1). The first term on the right-hand side of Eq. (6) represents the
′′

entropy production due to heat transfer irreversibility and invokes the rate of heat flow per
unit area and unit time, i.e., Fourier’s law

km Ttop + Tbot
(7)
��
Ṡ therm = (∇T)2 , with T̄ = .
T̄ 2 2
Ttop and Tbot are the temperature on top and bottom boundaries, respectively. Recall non-
dimensional variables Eq. (2)
T = T0 + (𝛥T)T ∗ ,
1
∇ = ∇∗ ,
d
𝛥T 2
(∇T)2 = 2 (∇∗ T ∗ )2 .
d
Substitute (∇T)2 in Eq. (7)

km 𝛥T 2 ∗ ∗ 2
(8)
��
Ṡ therm = (∇ T ) .
T̄ 2 d2
The second term on the right-hand side of Eq. (6) accounts for viscous dissipation effects
of the fluid
�� 𝜇f 2 𝜇f
Ṡ visc = 𝐪 + 𝛷, (9)
K T̄ T̄
where 𝐪 and 𝛷 are Darcy flux and the viscous dissipation function, respectively. The sec-
ond term on the right-hand side of Eq. (9) is only important when the flow tends to behave
like non-Darcy flow (Costa 2006) and can be neglected in our study. In general, the viscous
dissipation effects become increasingly important for a heterogeneous medium as preferred
fluid flow paths lead to a local increase in the fluid velocity. Take the scaling factor of 𝐪 in
Eq. (3) and plug in Eq. (9)
�� 𝛼𝛥Tgkm ∗ 2
Ṡ visc = (𝐪 ) .
̄ fd
Tc

We integrate entropy production over the computational domain, assuming homogeneous

material properties

13
202 P.-W. Huang, F. Wellmann

Fig. 3  Preferred cellular mode, (m, n), as a function of h1, h2. The figure is symmetric
√ with respect to the
line h1 = h2. The change between the modes (m, 0) and (m + 1, 0) occurs at h1 = m(m + 1) (Beck 1972).
We examine the convection pattern and entropy production of the plotted points using the proposed finite
element solver

Fig. 4  Temperature bound-

ary conditions and side length
(h1 , h2 ) of Beck’s box. The box
is closed with impermeable
boundaries

km (𝛥T)2 𝛼𝛥Tgkm
∫V T̄ 2 d2 ∫V ̄ f d ∫V
��
Ṡ total = Ṡ total dV = (∇∗ T ∗ )2 dV + (𝐪∗ )2 dV, (10)
Tc

where V is the volume of the computational domain. Dropping the asterisks, the dimen-
sionless entropy production, or the entropy generation number is

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 203

( )
d2 T̄ 2 Ṡ ̄
V ∫V 𝛥Tcf ∫V
1 2 𝛼dTg 2
NṠ = = (∇T) dV + (𝐪) dV . (11)
km (𝛥T)2 V

We would like to emphasize that the thermal entropy production measures the norm of the
linear and nonlinear parts of the temperature gradient, and the Nusselt number measures
the flux from nonlinear parts of the temperature from the top boundary. Since the entropy
production and the Nusselt number have different meanings in quality measure in convec-
tive heat transfer, we calculate both values for all of the simulations.

2.4 Problem Formulation

Börsing et al. (2017) investigated the entropy production of a naturally convecting porous
media of various aspect ratios and Rayleigh numbers in a 2D setting. We extended the
analysis to a 3D setting and designed numerical experiments over boxes of different dimen-
sions. Figure 3 shows a scatter plot of the box dimensions [h1 , h2 ] categorized into the tri-
angular test and the line test. The triangular test aims to inform the entropy production over
a wide range of cellular modes. The line test focuses on how entropy production relates to
both the total number of convection cells and the cellular modes’ dimensions. Throughout
the simulations, the Rayleigh number is set to 42.25, slightly above the critical Rayleigh
number. The characteristic length of the mesh is 0.05.
The initial condition of the transient problem is the conductive solution with a ±1%
perturbation. For each point in Fig. 3, we only realize one transient simulation. This is
certainly not ideal, as the steady-state solution of the transient problem depends strongly on
the initial condition. To compensate, we further analyze the line test by forcing steady-state
solutions using the initial conditions
� � � �
m𝜋x n𝜋z
T =A sin (𝜋y) cos cos + 1 − y,
h1 h2
√ � �−1 � � � �
m Ra m2 n2 m𝜋x n𝜋z
u=−A + 2 +1 cos (𝜋y) sin cos ,
h1 h21 h2 h1 h2
√
� �� �−1 � � � � (12)
m2 n2 m2 n2 m𝜋x n𝜋z
v =A Ra + 2 + 2 +1 sin (𝜋y) cos cos ,
h21 h2 h21 h2 h1 h2
√ � �−1 � � � �
n Ra m2 n2 m𝜋x n𝜋z
w=−A + + 1 cos (𝜋y) cos sin ,
h2 h21 h22 h1 h2

where A is the amplitude (Florio 2014). We set the amplitude such that the entropy produc-
tion does not exceed 2569
Ra, which is the analytical upper bound of the Nu–Ra relationship
(Doering and Constantin 1998).
The observations of varying Nusselt number with respect to 2D or 3D cellular modes
(Holst and Aziz 1972; Straus and Schubert 1978; Horne 1979) can be a reason for the
Nu–Ra discrepancy. We test this idea by simulating three boxes of different sizes that
have 2D or 3D preferred cellular modes during the onset of convection. The long boxes
[1.5, 1.0] and [2.3, 0.9] convect with a 2D cellular mode, and the wide boxes [2.5, 1.5] con-
vect with a 3D one. The three boxes are tested in the regions of 4𝜋 2 < Ra ≤ 196, and the
type of convection cells, the Nusselt number and the entropy production are reported. The

13
204 P.-W. Huang, F. Wellmann

characteristic length of the mesh is set to 0.1. The goal is to see how the transition of cellu-
lar modes due to increasing Rayleigh number influences the Nusselt number, and whether
it can explain the Nu–Ra discrepancy.

2.5 Basin Stability Analysis Using the Equivalent Entropy Production Initialization

Florio (2014) investigated the probability of convection modes in “critical boxes” using per-
turbation methods. The critical boxes are boxes with the size that lies on the transition between
several convection modes. We investigate this probability using the basin stability analysis
similar to Karani and Huber (2017a). Instead of exploring the amplitude space from -1 to 1,
we consider the symmetry of amplitudes, and we initialize the amplitudes in the space that is
bounded by a certain entropy production.
We pick one of the boxes Florio (2014) used in his analysis [21∕4 , 21∕4 ], approximated as
[1.189, 1.189]. The possible convection modes of this box when the Rayleigh √ number is criti-
cal are (0, 1), (1, 0) and (1, 1). The critical Rayleigh number is 𝜋 2 (2 + 3∕ 2) ≈ 40.68. We set
the Rayleigh number to 42.25 to be slightly above the critical Rayleigh number. We define the
initial condition and the amplitude of the modes (0, 1) and (1, 1)
( )
𝜋z
T01 = A01 sin (𝜋y) cos 1∕4 ,
(2 ) ( )
𝜋x 𝜋z
T11 = A11 sin (𝜋y) cos 1∕4 cos 1∕4 .
2 2
Apply the thermal part of nondimensional entropy production Eq. (11) to T01 and T11, and
sum the results
√ √
𝜋 2 ( 2 + 1) 2 𝜋 2 ( 2 + 2)
NSsum
̇ = NS01
̇ + NS11
̇ = √ (A01 ) + √ (A11 )2 .
4 2 8 2

Consider the amplitudes A01 and A11 as axes in the Cartesian coordinates, the combinations
of amplitudes that have the same amount of entropy production NSsum
̇ are an ellipse
√ √
(A01 )2 (A11 )2 4 2NSsuṁ 8 2NSsum
̇
+ = 1, a2 = √ , b2 = √ ,
a2 b2 2
𝜋 ( 2 + 1) 2
𝜋 ( 2 + 2)

where a is the semi-major axis and b is the semi-minor axis. We apply a change of variables
A01 = a cos 𝜉, A11 = b sin 𝜉,
( )
a
𝜉 = tan−1 tan 𝜃 ,
b
such that we can characterize the combination of amplitudes with the same amount of
entropy production using the angle 𝜃 , and the initial entropy production NSsum
̇ . The initial
conditions of velocities u, v, w are the same as Eq. (12). The probability of the (0, 1) mode
to occur is defined as the area that is characterized by 𝜃 = 0 to the separation angle 𝜃 = 𝜃 ∗
divided by quarter of the elliptical area
( ) 2𝜉
2 a
P(𝜃) = tan−1 tan 𝜃 = . (13)
𝜋 b 𝜋

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 205

We divide 𝜉 linearly into 17 parts from 0 to 90 degrees. The entropy productions NSsum
̇ we
choose to initialize are 1.02, 1.04, 1.08, 1.12 and 1.16. Note that we cannot perform such
analysis using the Nusselt number. However, the linear solutions of the convective tem-
perature Eq. (12) have zero contribution to the Nusselt number.

3 Results

3.1 Quality of Convective Heat Transfer in Different Settings of Boxes

Figure 5 shows the cellular mode and the entropy production of the line test with Ray-
leigh number of 42.25. Figure 6 represents the cellular mode and entropy production of
the triangular test with Ra = 42.25.

3.2 The Nu–Ra Relation

We test the Nusselt number of the boxes [1.5, 1.0], [2.3, 0.9] and [2.5, 1.5] with respect
to the Rayleigh numbers in the range between 42.25 and 196. Figures 7 and 8 are the
results.

3.3 The Probability of Mode Occurrence Using Basin Stability Analysis

Figure 9 shows the result of the basin stability analysis. We found out that the separa-
tion angle of the box [21∕4 , 21∕4 ] between modes (0, 1) and (1, 1) using the basin stabil-
ity analysis. The separation angle lies between 44.30◦ and 49.94◦. We average the two
angles and claim the separation angle is 47.12◦. Using Eq. (13), the probability of mode
(0, 1) to occur when the initial condition is the superposition of mode (0, 1) and (1, 1) is
46.8 percent. However, consider the symmetry of modes (0, 1) and (1, 0), we can calcu-
late the probability of mode (1, 1) to occur when the initial condition is the superposi-
tion of the three modes, which is 36.2 percent.

4 Discussion

4.1 Quality of Convective Heat Transfer in Different Settings of Boxes

Figure 5 shows that the cellular modes of our simulation match Beck’s prediction well.
The slight shift of the boundaries of cellular modes change can be attributed to that our
simulations are performed on a porous media with Ra = 42.25 > Rac . We color-coded
the cellular modes such that shades of blues and greens are for two-dimensional cellular
modes, and gradients of reds and yellows are for three-dimensional cellular modes. By
only looking at the entropy production of the blues and greens, the patterns are similar
to Börsing et al. (2017)’s tests of 2D boxes. It is clear that box sizes are relevant to the
quality of convective heat transfer. Focusing on the whole figure including the three-
dimensional cellular modes, the observations are:

13
206 P.-W. Huang, F. Wellmann

Fig. 5  The line test at Ra=42.25 plotted over Beck’s diagram (Beck 1972). The y-axis of the middle and
bottom plots represents the entropy production. The middle plot is solved by the transient solver with the
randomly perturbed initial conditions. The bottom plot is solved by the steady-state solver with the initial
conditions Eq. (12). The bracketed values (m, n) represent the cellular mode

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 207

Fig. 6  The triangular test at Ra=42.25 plotted over Beck’s diagram (Beck 1972). The bottom plot repre-
sents the contour of entropy production using linear interpolation. The bracketed values (m, n) represent the
cellular mode

1. The quality of convective heat transfer of three-dimensional cellular modes is generally

worse than those of two-dimensional cellular modes when Ra= 42.25.
2. The entropy production gradually increases from three dimensional cellular modes to
two-dimensional cellular modes that have the same total sum of cells m + n.

13
208 P.-W. Huang, F. Wellmann

Fig. 7  The relation between the Rayleigh number and the entropy production of the boxes [1.5, 1.0],
[2.3, 0.9] and [2.5, 1.5]. The bracketed values (m, n) represent the cellular mode

3. The transient solution does not necessary converge to cellular modes with better quality
of convective heat transfer. See Sect. 4.3 for further discussions about the influence of
the initial conditions.

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 209

Fig. 8  The Nu–Ra relation of the boxes [1.5, 1.0], [2.3, 0.9] and [2.5, 1.5], plotted over Cheng (1979)’s
compilation

Fig. 9  The diagram of basin

stability analysis. The red line
is drawn as y = tan (𝜃 ∗ x), where
𝜃 ∗ is the separation angle. The
streamlines show how differ-
ent initial conditions reach the
steady-state solution. The brack-
eted values (m, n) represent the
cellular mode at steady state

13
210 P.-W. Huang, F. Wellmann

Figure 6 further supports the aforementioned observations. On the boundaries from

mode (2, 0) to either (2, 1) or (1, 1), we can also observe a descent of the entropy pro-
duction. Increasing convection cells do not guarantee a better quality of convective heat
transfer. These results give us new insights into the relation between cellular modes and
entropy production. We can also view these results as more rigorous benchmarks that
compare with theoretical predictions.

4.2 The Nu–Ra Relation

Figure 8 compared the Nu–Ra relation of the three boxes with previous experimental and
numerical results. The results do not show the wide scattering of Nusselt number during the
onset of convection nor in the region of Ra ≤ 100. This can be attributed to the lack of thermal
dispersion effects in our model (Karani et al. 2017).
However, the regions of 140 <Ra< 196 show a wide scattering of the Nusselt number
from 3.212 to 4.145, which can be explained by the multiple steady states of convection pat-
tern. Our results agree with Straus and Schubert (1979)’s numerical tests that both two and
three-dimensional convection cells could be obtained at 60 ≤Ra≤ 150 by perturbing the ini-
tial condition of temperature. The results also agree with Borkowska-Pawlak and Kordylewski
(1985)’s proof, that continuous transition of pattern flows from two-dimensional to three-
dimensional structure is possible (and vice-versa) with Rayleigh number variations.
The hypotheses of box sizes and multiple steady states of convection pattern influencing
the Nu–Ra relation are proved using the three boxes experiment. Future experiments of the
Nu–Ra relation should also be aware of the chosen box size and the convection pattern of
fluids.

4.3 The Probability of Mode Occurrence and its Implications

Our results suggest that we cannot infer the steady-state pattern, given the entropy pro-
duction or the Nusselt number of the initial condition. The steady-state solution is deter-
mined by the initial combination of mode amplitudes. Thus, we investigate the probability
of mode occurrence, assuming equal chance for amplitude combinations to occur as initial
conditions.
Using the basin stability analysis with the equivalent entropy production initialization,
we can calculate the probability of mode occurrence. In our example of the [21∕4 , 21∕4 ] box,
the probability of mode (1, 1) to occur is 36.2 percent, which is slightly higher than the
modes (0, 1) and (1, 0). From the simulation, we also know that the Nusselt number for
mode (0, 1) and (1, 1) is 1.0637 and 1.0546, respectively. Combining with their occur-
ring probability, we can, therefore, calculate the expectation value for the Nusselt number,
which is 1.0579. We have provided a method for a new Nu–Ra relationship in the non-
oscillating region of the Rayleigh number.
The assumption that we can use a straight line to separate the probability space is only
for convenience. It is also possible that the line can be of a higher order. As in Fig. 9, the
straight line does not separate the modes. Nevertheless, in this example, it already gives a
good approximation and demonstrates how we can apply this method to different box sizes
and Rayleigh numbers.

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 211

5 Conclusion

We show how the influence of different box sizes and multiple steady states of convec-
tion pattern leads to the discrepancy of Nu–Ra relation in the region of moderate Ray-
leigh number. We also demonstrate the method of basin stability analysis using equivalent
entropy production initialization to study the probability of mode occurrences in naturally
convected porous media. This method can be utilized for further studies of the Nu–Ra
relationship.

Appendix: Methods of Cellular Mode Checking

In the appendix section, we discuss the methods we used for checking cellular modes of a
certain convection pattern.

Counting by Visual Inspection

We can count how many cells a convection pattern has by looking at the temperature pro-
file at y = 0.5. The method is: count how many times of a temperature change and a 0.5 (-)
temperature, in either x or z-direction. See Fig. 10 for better comprehension of this concept.
This process is programmed and used to post-process the simulation results in Fig. 5, 6, 7
and 8. The problem with this method is that it cannot differentiate between the 3D cellular
modes (bottom left of Fig. 10) and the linear combination of the 2D cellular modes (bot-
tom right of Fig. 10).

Fig. 10  Examples of automatic cellular mode checking using different configurations of boxes. The title of
each plot represents the cellular mode (m, n)

13
212 P.-W. Huang, F. Wellmann

Counting by Fourier Analysis

Our finite element code calculates the temperature in space T(x, y, z) for a given time. The
numerical solution consists of the linear solutions Eq. (12); therefore, we can apply the
concept of Fourier analysis. We define the convection temperature as the full solution with-
out the conduction solution
Tconvect =T(x, y, z) − (1 − y)
∞
∑
= Amn Tmn ,
m,n=0

note that m and n cannot be 0 at the same time. If we want to know if the system is convect-
ing by a certain cellular mode (m = M, n = N), we multiply the convection temperature by
the linear solution TMN and integrate through the domain 𝛺
∞
∑
∫𝛺 ∫𝛺
TMN Tconvect d𝛺 = Amn TMN Tmn d𝛺
m,n=0

∫𝛺
=AMN (TMN )2 d𝛺.

The amplitude AMN is therefore

∫𝛺 TMN Tconvect d𝛺
∫𝛺 (TMN )2 d𝛺
AMN = .

If M ≠ 0 and N ≠ 0,

h1 h2 ∫𝛺 MN convect
8
AMN = T T d𝛺.

If either of M or N is 0,

h1 h2 ∫𝛺 MN convect
4
AMN = T T d𝛺.

This Fourier analysis not only gives us the information of which cellular mode the system
has, but also the value of amplitudes. The value of amplitudes can also be used to differen-
tiate the 3D cellular modes from the linear combination of the 2D cellular modes. We use
this method to analyze Fig. 9.

Acknowledgements The authors gratefully acknowledge the computing time granted through JARA-HPC
on the supercomputer JURECA at Forschungszentrum Jülich. Since the first author has been studying in
the Geothermal Energy and Geofluids group from August 2018, he thanks the Werner Siemens Foundation
for their endowment of the Geothermal Energy and Geofluids Group at the Institute of Geophysics, ETH
Zürich. We thank the reviewers for their positive comments, which improved the manuscript.

Funding Open Access funding provided by Lib4RI – Library for the Research Institutes within the ETH
Domain: Eawag, Empa, PSI &

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as
you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons

13
An Explanation to the Nusselt–Rayleigh Discrepancy in Naturally… 213

licence, and indicate if changes were made. The images or other third party material in this article are
included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the mate-
rial. If material is not included in the article’s Creative Commons licence and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit [Link]

References
Beck, J.L.: Convection in a box of porous material saturated with fluid. Phys. Fluids. 15(8), 1377–1383
(1972). https​://[Link]/10.1063/1.16940​96
Bejan, A.: Convection Heat Transfer. Wiley, Hoboken (2013)
Borkowska-Pawlak, B., Kordylewski, W.: Cell-pattern sensitivity to box configuration in a saturated porous
medium. J. Fluid. Mech. 150, 169–181 (1985). https​://[Link]/10.1017/S0022​11208​50000​88
Börsing, N., Wellmann, J.F., Niederau, J., Regenauer-Lieb, K.: Entropy production in a box: analysis of
instabilities in confined hydrothermal systems. Water. Resour. Res. 53, 7716–7739 (2017). https​://doi.
org/10.1002/2017W​R0204​27
Brunner, P., Simmons, C.T.: HydroGeoSphere: a fully integrated. Phys. Based Hydrol. Model. Groundwater
50(2), 170–176 (2012). https​://[Link]/10.1111/j.1745-6584.2011.00882​.x
Cheng, P.: Heat transfer in geothermal systems. Adv. Heat Trans. 14, 1–105 (1979). https​://[Link]/10.1016/
S0065​-2717(08)70085​-6
Costa, V.A.F.: On natural convection in enclosures filled with fluid-saturated porous media including vis-
cous dissipation. Int. J. Heat. Mass. Tran. 49(13–14), 2215–2226 (2006). https​://[Link]/10.1016/[Link]​
tmass​trans​fer.2005.12.011
Doering, C.R., Constantin, P.: Bounds for heat transport in a porous layer. J. Fluid. Mech. 376, 263–296
(1998). https​://[Link]/10.1017/S0022​11209​80028​1X
Diersch, H.J.: FEFLOW: Finite Element Modeling of Flow, Mass and Heat Transport in Porous and Frac-
tured Media. Springer-Verlag, Berlin, Heidelberg (2014). https​://[Link]/10.1007/978-3-642-38739​-5
Elder, J.W.: Transient convection in a porous medium. J. Fluid. Mech. 27(3), 609–623 (1967). https​://doi.
org/10.1017/S0022​11206​70005​76
Florio, B.J.: The interaction of convection modes in a box of a saturated porous medium. J. Eng. Math.
86(1), 71–88 (2014). https​://[Link]/10.1007/s1066​5-013-9647-4
Gaston, D., Newman, C., Hansen, G., Lebrun-Grandié, D.: MOOSE: A parallel computational framework
for coupled systems of nonlinear equations. Nucl. Eng. Des. 239(10), 1768–1778 (2009). https​://doi.
org/10.1016/[Link]​gdes.2009.05.021
Govorukhin, V.N., Shevchenko, I.V.: Multiple equilibria, bifurcations and selection scenarios in cosymmet-
ric problem of thermal convection in porous medium. Physica. D 361(15), 42–58 (2017). https​://doi.
org/10.1016/[Link]​.2017.08.012
Graf, T., Boufadel, M.C.: Effect of viscosity, capillarity and grid spacing on thermal variable-density flow. J.
Hydrol. 400(1–2), 41–57 (2011). https​://[Link]/10.1016/[Link]​ol.2011.01.025
Herwig, H.: What exactly is the Nusselt number in convective heat transfer problems and are there alterna-
tives? Entropy 18(5), 198–212 (2016). https​://[Link]/10.3390/e1805​0198
Hewitt, D.R., Neufeld, J.A., Lister, J.R.: High Rayleigh number convection in a three-dimensional porous
medium. J. Fluid. Mech. 748, 879–895 (2014). https​://[Link]/10.1017/jfm.2014.216
Holst, P.H., Aziz, K.: Transient three-dimensional natural convection in confined porous media. Int. J. Heat.
Mass. Trans. 15(1), 73–90 (1972). https​://[Link]/10.1016/0017-9310(72)90167​-6
Horne, R.N.: Three-dimensional natural convection in a confined porous medium heated from below. J.
Fluid. Mech. 92(4), 751–766 (1979). https​://[Link]/10.1017/S0022​11207​90008​60
Karani, H., Huber, C.: Transitional behavior of convective patterns in free convection in porous media. J.
Fluid. Mech. 818, R3 (2017a). https​://[Link]/10.1017/jfm.2017.155
Karani, H., Huber, C.: Role of thermal disequilibrium on natural convection in porous media: insights from
pore-scale study. Phys. Rev. E. 95, 033123 (2017b). https​://[Link]/10.1103/PhysR​evE.95.03312​3
Karani, H., Rashtbehesht, M., Huber, C., Magin, R.L.: Onset of fractional-order thermal convection in
porous media. Phys. Rev. E. 96, 063105 (2017). https​://[Link]/10.1103/PhysR​evE.96.06310​5
Lister, C.R.B.: An explanation for the multivalued heat transport found experimentally for convection in a
porous medium. J. Fluid. Mech. 214, 287–320 https​://[Link]/10.1017/S0022​11209​00001​43
Menck, P.J., Heitzig, J., Marwan, N., Kurths, J.: How basin stability complements the linear-stability para-
digm. Nat. Phys. 9, 89–92 (2013). https​://[Link]/10.1038/nphys​2516

13
214 P.-W. Huang, F. Wellmann

Simmons, C.T., Elder, J.W.: The elder problem. Groundwater 55, 926–930 (2017). https​://[Link]/10.1111/
gwat.12593​
Straus, J.M., Schubert, G.: On the existence of three-dimensional convection in a rectangular box containing
fluid-saturated porous material. J. Fluid. Mech. 87(2), 385–394 (1978). https​://[Link]/10.1017/S0022​
11207​80016​64
Straus, J.M., Schubert, G.: Three-dimensional convection in a cubic box of fluid-saturated porous material.
J. Fluid. Mech. 91(1), 155–165 (1979). https​://[Link]/10.1017/S0022​11207​90000​82
Thiesen, M., Scheel, K., Diesselhorst, H.: Untersuchungen über die thermische Ausdehnung von festen und
tropfbar flüssigen Körpern. Wiss. Abhandlungen der Physik.-Techn. Reichsanst. 3, 1–70 (1900)
Trefry, M.G., Muffels, C.: FEFLOW: a finite-element ground water flow and transport modeling tool.
Groundwater 45(5), 525–528 (2007). https​://[Link]/10.1111/j.1745-6584.2007.00358​.x
Vadasz, P.: Heat flux dispersion in natural convection in porous media. Int. J. Heat. Mass. Trans. 53(17–18),
3394–3404 (2010). https​://[Link]/10.1016/[Link]​tmass​trans​fer.2010.03.034
Vadasz, P., Braester, C.: The effect of imperfectly insulated sidewalls on natural convection in porous media.
Acta. Mech. 91(3–4), 215–233 (1992). https​://[Link]/10.1007/BF011​94111​
Voss, C.I.: A finite-element simulation model for saturated-unsaturated, fluid-density-dependent ground-
water flow with energy-transport or chemically-reactive single-species solute transport. Water Resour.
Invest. Rep. (1984). https​://[Link]/10.3133/wri84​4369
Yudovich, V.I.: Cosymmetry, degeneration of solutions of operator equations, and onset of a filtration con-
vection. Math. Notes Acad. Sci. USSR 49(5), 540–545 (1991). https​://[Link]/10.1007/BF011​42654​

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

13