Tribology International 124 (2018) 247–258

Contents lists available at ScienceDirect

Tribology International
journal homepage: www.elsevier.com/locate/triboint

An effective Navier-Stokes model for the simulation of textured surface T

lubrication
Michael Rom∗, Siegfried Müller
Institut für Geometrie und Praktische Mathematik (IGPM), RWTH Aachen University, Templergraben 55, 52056, Aachen, Germany

A R T I C LE I N FO A B S T R A C T

Keywords: For a wide range of applications in the context of textured surface lubrication, both the Reynolds equation and
Surface texture the Stokes system are not valid. When it comes to flows with high velocities and surfaces with deep textures of
Hydrodynamic lubrication moderate length, the Navier-Stokes equations have to be solved instead. For this kind of applications, a new
Numerical analysis effective model is presented which is derived from an upscaling technique applied to the incompressible Navier-
Journal bearing
Stokes equations. It provides more accurate results than the Reynolds or Stokes problem. In addition, the
computational cost compared with solving the Navier-Stokes system is reduced significantly. Numerical results
for two- and three-dimensional test cases demonstrate the accuracy and efficiency of the approach.

1. Introduction
Whenever tribological optimization of surfaces is investigated, tex-
tured surfaces can be considered to improve the hydrodynamic per-
formance, for instance by reducing friction. Many experimental and
theoretical works have demonstrated the potential of texturing, e.g.,
[1–5]. Since the microscale effects induced by the textures are not fully
understood yet, textured surface lubrication is an open research field
[6]. For a particular application, it is usually not clear in advance how
to choose the texturing for an optimal result. There is a large number of
parameters influencing the fluid behavior, e.g., the geometry (shape,
length, width, height) or the distribution of the textures but also the
operating conditions. Hence, numerical investigations usually require a
large number of simulations, often with highly resolved computational
meshes, leading to the demand of fast and accurate solution schemes.
The Reynolds lubrication equation [7] can be used for the numerical
investigation of fluid film lubrication between two surfaces. It is de-
rived from the Navier-Stokes equations by assuming a negligible pres-
sure variation over the small fluid film thickness and by neglecting
inertia forces. These assumptions lead to a reduction of the spatial di-
mension by one. Furthermore, the resulting Reynolds equation is linear.
Hence, solving the Reynolds equation is much less expensive than sol-
ving the nonlinear Navier-Stokes equations. When rough or textured
surfaces are considered, a further reduction of the computational effort
can be achieved by applying averaging techniques such as the average
flow model by Patir and Cheng [8] or the mathematical concept of
homogenization, see for instance [9,10], instead of resolving the mi-
croscale effects with very fine computational meshes. In the approach
by Patir and Cheng, shear and pressure flow factors are computed on
the microscale, characterizing the differences between the flow beha-
vior of rough and smooth surfaces. Subsequently, the macroscale be-
havior can be resolved using these flow factors. The homogenization
approach, which is similar, in addition allows for capturing microscale
effects. However, it was shown in numerous studies that the Reynolds
equation and, hence, also the averaged models are not valid for many
relevant applications involving textured surfaces, see for instance
[11–13]. In particular, deep textures lead to recirculation of the flow in
the textures, and high flow velocities induce inertia effects. Both phe-
nomena cannot be captured by the Reynolds equation, resulting in the
computation of inaccurate pressure distributions. The Stokes system is
recommended for applications with deep textures but also fails for high
flow velocities [6].
Based on the average flow model by Patir and Cheng, many ex-
tensions were implemented, especially to account for surface contact
[14] or cavitation [15]. However, approaches for improving the va-
lidity regarding recirculation or inertia effects are rare: de Kraker et al.
presented a texture averaged Reynolds equation [16,17] for which they
modified the approach by Patir and Cheng by solving the Navier-Stokes
equations instead of the Reynolds equation on the microscale. This al-
lows for capturing recirculation and inertia effects on the microscale
and leads to an adjustment of the flow factors for the macroscale
computation. To the authors' knowledge, there has never been an
evaluation of the method by a full simulation study to find out whether
a substantial improvement of the validity compared with the average
flow model by Patir and Cheng can be obtained. Instead, de Kraker et al.
only compared their resulting flow factors with those of the average

∗
Corresponding author.
E-mail address: rom@igpm.rwth-aachen.de (M. Rom).

https://doi.org/10.1016/j.triboint.2018.04.011
Received 7 February 2018; Received in revised form 6 April 2018; Accepted 10 April 2018
Available online 14 April 2018
0301-679X/ © 2018 Elsevier Ltd. All rights reserved.
M. Rom, S. Müller Tribology International 124 (2018) 247–258

Nomenclature ht texture height

lt texture length
α gradient angle due to eccentricity n normal vector
ε roughness scale nt number of textures
η dynamic viscosity p pressure
θp upscaling error regarding pressure p0 zeroth-order pressure (macroscale)
θu upscaling error vector regarding velocity p1 cell problem pressure (microscale)
ν kinematic viscosity Q volumetric flow rate
ρ density R revolutions
τ shear stress Re Reynolds number
ψ relative clearance t parameter for B-spline curve
Γ boundary of computational domain u velocity vector
Γ1, …, Γ4 left, right, lower, upper domain boundary u velocity component in flow direction
Γeff effective domain boundary in effective problem u0 zeroth-order velocity vector (macroscale)
Γ,l Γr left, right domain boundary in cell problem u1 cell problem velocity vector (microscale)
Γper periodic domain boundary u1, i velocity solution vector of ith cell problem
Γt texture domain boundary in cell problem uRe Reynolds velocity vector
Γup upper domain boundary in cell problem u rot rotational velocity vector
Ω computational domain v velocity component in film thickness direction
Ω0 computational domain for zeroth-order problem w velocity component in width direction (only 3D)
Ωcp computational domain for cell problem W load-carrying capacity
Ωeff computational domain for effective problem wb bearing width (only 3D)
ΩRe computational domain for Reynolds problem wt texture width (only 3D)
c radial clearance x macroscale coordinate vector
d bearing diameter x macroscale coordinate in flow direction
dt , x distance between textures in flow direction xp reference point
dt , z distance betw. textures in width dir. (only 3D) y microscale coordinate vector
e absolute eccentricity y macroscale coordinate in film thickness direction
erel relative eccentricity z macroscale coordinate in width dir. (only 3D)
F friction force
h film thickness

flow model. for the smooth Navier-Stokes problem in step (iii) .

As a consequence of the shortcomings mentioned above, one has to
solve the Navier-Stokes equations when the Reynolds and Stokes solu-
tions are inaccurate. Due to the high computational effort, this can
usually only be done for two-dimensional setups or for configurations
with a single or a few textures [2,11,18,19]. Hence, there is a demand
for a method which is faster than solving the Navier-Stokes equations
and provides more accurate solutions than the Reynolds equation or the
Stokes system.
In this paper, we present a new approach based on an upscaling
strategy applied to the incompressible Navier-Stokes equations. The
general theory which has been developed for compressible flows over
surfaces with a periodic roughness is described in Ref. [20]. It can be
understood as a three-step process: (i) solving the Navier-Stokes equa-
tions in a smooth domain which replaces the original rough domain,
resulting in the zeroth-order solution, (ii) solving linear microscale cell
problems for each roughness element to compute a roughness-induced
correction for the zeroth-order solution and (iii) solving the Navier-
Stokes equations in a smooth domain with effective boundary conditions
computed from the zeroth-order and the cell problem solutions. This
upscaling approach only works for roughness scales being small com-
pared with the dimensions of the flow. For textured surface lubrication,
this requirement is often violated because the fluid film thickness can
even be smaller than the texture height. Hence, we propose modifica-
tions to the upscaling approach for textured surface lubrication in this
paper: instead of obtaining the zeroth-order solution from the Navier-
Stokes equations in a smooth domain, see step (i) , we solve the Rey-
nolds equation for the original textured problem. In step (ii) , the re-
sulting Reynolds velocity is used instead of the zeroth-order solution in
the cell problems. Away from the textures, it is a good approximation of
the Navier-Stokes velocity, even for surfaces with deep textures. The
cell problem solutions then provide the effective boundary conditions
Our new effective approach is suitable for two- and three-dimen-
sional computations. To validate this, we present numerical simulation
results for different 2D and 3D setups of textured journal bearings.
These results are compared qualitatively and quantitatively with
Navier-Stokes, Reynolds and Stokes solutions. In addition, the compu-
tational cost for solving the effective problem is measured and related
to the cost for solving the Navier-Stokes system to evaluate the speed-
ups obtained by the effective method.
Note that up to now we have not incorporated a cavitation model
into our effective approach since our focus has been on the fundamental
feasibility of our scheme.
The present paper is structured as follows: in Sect. 2, our effective
Navier-Stokes model is derived from the general upscaling approach.
The new method is evaluated in Sect. 3 by studying simulation results
and computational cost. The conclusion in Sect. 4 summarizes our
findings and discusses potential improvements of the effective model.
2. Derivation of the effective Navier-Stokes model
For the derivation of our effective Navier-Stokes model, we focus on
textured journal bearing investigations to have an exemplary applica-
tion at hand. We emphasize that the approach is not limited to this case
and other applications in the context of textured surface lubrication are
conceivable. The computational domain is schematically depicted in
Fig. 1, for the ease of illustration for a two-dimensional setting. Note
that the journal bearing is unfolded. In the three-dimensional case,
additional boundary conditions are prescribed on the sides of the par-
ticular domain as detailed in the following where necessary. The film
thickness without taking into account the textures is given by

248
M. Rom, S. Müller Tribology International 124 (2018) 247–258

∇y ⋅u1 = 0 ,
1
(u 0⋅∇y ) u1 + ∇y p1 − Δ u
εRe y 1
=0. (6)
On the lower boundary Γt including the texture, the no-slip condi-
tion
Fig. 1. Two-dimensional computational domain Ω with boundaries Γ1, …, Γ4 for u 0 (xp + εy ) + εu1 (y ) = 0 , y ∈ Γt (7)
the incompressible Navier-Stokes equations.
is imposed. On the upper boundary Γup , the natural condition
2x ⎞ 1
h (x ) = c ⎛1 + erel cos ⎛ ⎞ , (∇y u1) n − p1 n = 0 on Γup
(8)
⎝ ⎝ d ⎠⎠ (1) εRe

where c, d and erel = e / c are the radial clearance, the bearing diameter
and the relative eccentricity (eccentricity related to clearance), re-
spectively, and x denotes the coordinate in flow direction. In this ex-
ample, the textures have a rectangular shape. Consequently, the texture
height ht is constant and has to be added to the film thickness h (x )
where applicable. The rotational velocity of the shaft surface is u = u rot
with the velocity vector given by u = (u, v )T or u = (u, v, w )T in the
two- or three-dimensional case, respectively.
The starting point for the derivation of our effective model is the
dimensionless stationary incompressible Navier-Stokes system
∇⋅u = 0,
1 u (x ) = u 0 (x ) + ε u1, i ,
(u⋅∇) u + ∇p − Re
Δu =0 (2)
with velocity vector u , pressure p and Reynolds number Re = urot c / ν .
Here, urot is the velocity component of the shaft in rotational direction
and ν the constant kinematic viscosity of the lubricant. On the bound-
aries Γ1 and Γ2 , we use the natural boundary condition
1 u1, i : =
(∇u ) n − pn = 0 on Γ1, Γ2,
Re (3)
where n is the outward-pointing normal. This condition is also used for
the sides of the bearing for three-dimensional computations. On Γ3 and
Γ4 , we impose the Dirichlet conditions
u = 0 on Γ3, (4a)
u = u rot on Γ4 . (4b)
is used. On the boundaries Γper , in 3D including the sides of the domain,
u1 and p1 are prescribed to be periodic. The solution (u1, p1 ) of the
system (6), (7), (8) can be understood as a scaled correction of the
macroscale solution (u 0, p0 ) .
3. Effective problem (macroscale): The domain Ωeff for the effective
problem is smooth. In this domain, the incompressible Navier-Stokes
system (2) is solved with boundary conditions (3) and (4b), where
(3) is also used on the sides of the bearing in case of a three-di-
mensional computation. The no-slip condition (4a) for the lower
boundary is replaced by the effective boundary condition
x ∈ Γeff (9)
such that the zeroth-order solution u 0 and the cell problem solutions u1, i
enter the effective problem. For each cell problem i, its corresponding
mean value u1, i is computed along the respective upper cell problem
boundary Γup , i.e.,
1
Γup
∫Γ
up
u1, i (y ) dγ .
(10)

2.1. Upscaling strategy for the Navier-Stokes equations

In the following, we briefly describe the upscaling approach applied

to the incompressible Navier-Stokes system (2). Details can be found in
Ref. [20]. The process is illustrated in Fig. 2. It is based on the
asymptotic expansion of the solution of the original problem which
provides
u = u 0 + εu1 + ε 2θu ,
p = p0 + εp1 + ε 2θp, (5)
where (u 0, p0 ) is the zeroth-order solution on the macroscale, (u1, p1 ) is
the solution of the cell problem on the microscale, ε denotes the
roughness scale, cf. Fig. 2, and (θu , θp) is the unknown upscaling error
that is assumed to be smooth.
The upscaling process consists of the following three steps:

1. Zeroth-order problem (macroscale): The incompressible Navier-

Stokes system (2) with boundary conditions (3) and (4) is solved in
the domain Ω0 which is smooth and encloses the roughness or in our
case the textures.
2. Cell problems (microscale): The cell problems are solved in the
local domain Ωcp with the local coordinate y = (x − xp)/ ε , where xp
is a reference point, cf. Fig. 2. The number of cell problems to be
solved is the same as the number of textures. The system of equa-
tions is similar to the Navier-Stokes equations but linear. This is due
to the zeroth-order velocity solution u 0 entering the momentum Fig. 2. Upscaling approach: the computation in the original textured domain Ω
equation such that is replaced by a three-step solution process in the domains Ω0 , Ωcp and Ωeff .

249
M. Rom, S. Müller Tribology International 124 (2018) 247–258

The upscaling approach as presented in this section was developed

for rough rather than textured surfaces. Hence, the roughness scale ε is
usually much smaller than the dimensions of the flow. Due to small film
thicknesses, this requirement is often violated for textured surface lu-
brication and the approach cannot be applied. We conducted numerical
simulations which confirmed that only test cases with very shallow
textures led to reasonable results. The main problem for deeper textures
lies in the zeroth-order velocity u 0 which is too inaccurate such that the
cell problem solutions u1 fail to correct it appropriately.
In search of a more accurate velocity u 0 , we investigated the
Reynolds velocity uRe . How this can be computed from the solution of
the Reynolds equation, is briefly derived in Sect. 2.2.

2.2. Computation of Reynolds velocity

For deriving the stationary incompressible Reynolds equation

h3 1
− ∇⋅⎛Re ∇p ⎞ + u rot⋅∇h = 0,
⎜ ⎟

⎝ 12 ⎠ 2 (11)
the starting point is the three-dimensional Navier-Stokes system (2):
neglecting inertia forces, particular velocity gradients and pressure
changes in film thickness direction y leads to the system
∂p 1 ∂ 2u ∂p ∂p 1 ∂ 2w
= , = 0, = .
∂x Re ∂y 2 ∂y ∂z Re ∂y 2 (12)
For the first of these three equations, two integrations with respect
to the film thickness direction y yield Fig. 3. Modified effective model: as in the upscaling approach, the computation
in the original textured domain Ω is replaced by a three-step solution process.
The main difference is the replacement of the zeroth-order problem (Ω0 ) with
1 2 ∂p 1 ⎛ ∂u ⎞
y = ⎜u − u −y ⎟. the Reynolds problem which reduces the spatial dimension (ΩRe ).
2 ∂x Re ⎜ ∂y y=0 ⎟
⎝ y=0 ⎠ (13)
Note that to simplify the derivation the film thickness direction y is solved such that (14), (15) and (17) can be computed from the pressure
reversed here compared with the illustration in Fig. 1: the rotating shaft gradient. Note that in our setting wrot = 0 and ∂h/ ∂z = 0 such that (15)
surface is at y = 0 , and the bearing surface is at y = h and curved due to and (17) are simplified.
the eccentricity. With the no-slip conditions u = u rot for y = 0 and Our investigations have shown that the Reynolds velocity (18)
u = 0 for y = h , the constants in (13) can be determined such that computed for a textured surface is a good approximation of the velocity
resulting from the Navier-Stokes equations, at least away from the
1 ∂p y textures. This is demonstrated in Sect. 3.2. How the Reynolds velocity is
u= Re y (y − h) + urot ⎛1 − ⎞ .
2 ∂x ⎝ h⎠ (14) incorporated into our effective Navier-Stokes model, is presented in
An analogous derivation for the third equation in (12) results in Sect. 2.3.

1 ∂p y
w= Re y (y − h) + wrot ⎛1 − ⎞ .
2 ∂z ⎝ h⎠ (15) 2.3. Effective Navier-Stokes model based on the Reynolds equation
For more details on this derivation, see for instance [21].
The velocity component v in film thickness direction y can be de- Our modifications of the upscaling approach are illustrated in Fig. 3.
termined by integrating the continuity equation, leading to The main difference is the replacement of the zeroth-order problem
with the Reynolds problem, i.e., in the first step of the three-step pro-
y ∂u y ∂w
v=− ∫0 ∂x
dy− ∫0 ∂z
dy .
(16)
cess the Reynolds equation has to be solved for the original textured
problem. This reduces the spatial dimension, cf. domain ΩRe in Fig. 3.
From (14) and (15), the derivatives ∂u/ ∂x and ∂w / ∂z can be com- Further changes concern the domain and the boundary conditions of
puted. By using (11), second-order derivatives of the pressure can be the cell problem and the lower boundary conditions of the effective
eliminated such that the velocity component v is given by problem. Our new effective Navier-Stokes model is described in detail
below. Note that we do not distinguish between macroscale and mi-
∂h ∂h y 2 y 1 ∂p ∂h ∂p ∂h ⎞ 2 y
v = ⎛urot + wrot ⎞ 2 ⎛1 − ⎞ + Re ⎛ + y ⎛ − 1⎞ . croscale anymore.
⎝ ∂x ∂z ⎠ h ⎝ h ⎠ 2 ⎝ ∂x ∂x ∂z ∂z ⎠ ⎝ h ⎠
(17) 1. Reynolds problem: The stationary incompressible Reynolds equa-
After reversing the film thickness direction y again to obtain the tion (11) is solved for the original textured problem with periodic
velocities in our coordinate system depicted in Fig. 1, the components boundary conditions for the pressure on Γper . Zero mean pressure is
of the Reynolds velocity vector uRe can finally be computed by prescribed to obtain a unique solution. From the resulting pressure
gradient, the Reynolds velocity (18) is computed as described in
uRe = u cosα, v Re = u sinα − v, w Re = w , (18)
Sect. 2.2. It replaces the zeroth-order velocity solution u 0 .
where the angle α is defined by
2 2x In a three-dimensional setting, for which the Reynolds problem due
tanα = h′ (x ) = − c erel sin ⎛ ⎞. to the reduction of the spatial dimension is two-dimensional, we pre-
d ⎝ d ⎠ (19)
scribe p = 0 on the sides of the bearing.
For the calculation of (18), the Reynolds equation (11) has to be

250
M. Rom, S. Müller Tribology International 124 (2018) 247–258

2. Cell problems: In the cell problem domain Ωcp , we do not use the ratios of texture length to texture height, the Reynolds equation and the
local coordinate y anymore. The system of equations is similar to (6) Stokes system are not valid as investigated in [13] and shown in Sect.
and reads 3.3. Information on the 2D simulation setups such as the chosen dis-
cretizations or solvers are given in Sect. 3.1. In Sect. 3.2, the quality of
∇⋅u1 = 0 ,
1
the Reynolds velocity (18) for approximating the velocity obtained by
(uRe⋅∇) u1 + ∇p1 − Re
Δu1 =0. (20) solving the Navier-Stokes simulations is investigated. Simulation results
for the 2D test cases are presented in Sect. 3.3. A 3D setup is described
The domain has been changed such that in the case of rectangular
in Sect. 3.4, and the applicability of our concept to this setup is de-
textures the sharp corners are avoided. This simplifies mesh generation
monstrated in Sect. 3.5. By comparing computation times of Navier-
for the cell problems and, more importantly, eliminates pressure sin-
Stokes and effective simulations, the performance of our new approach
gularities from the simulation. The texture boundary Γ3 still includes the
is studied in Sect. 3.6.
whole texture, i.e., all parts of the boundary of the domain below the
dashed line, cf. Fig. 3. We call it Γ3 instead of Γt because the no-slip
condition (4a), u1 = 0 , is used on this boundary. Another modification 3.1. 2D simulation setups
of the cell problem domain concerns the upper boundary: Γup is replaced
by Γ4 which is adapted to the original boundary by considering the The simulation parameters regarding the textured journal bearing,
eccentricity. Hence, the shape of the domain depends on the coordinate the lubricant and the operating conditions are summarized in Table 2.
x and the mesh has to be modified for each cell problem. This can be Since the eccentricity of the system has a large influence on the pressure
done automatically by applying a mesh transformation function which distribution, we decided to vary it and set up two test cases with
distributes the mesh points uniformly in film thickness direction y ac- erel = 0.3 and erel = 0.8, both with nt = 15 textures. To avoid any pres-
cording to the film thickness at the particular position x. The boundary sure singularities, we used a smooth texture shape described by a cubic
condition on Γ4 is given by (4b), u1 = u rot . On the boundaries Γl and Γr , B-spline curve with uniform knot vector (0 , 0, 0, 0, 0.25, 0.5, 0.75, 1, 1,
we prescribe 1, 1) and control points (0,0) , (0.2 lt , 0) , (0.2 lt , −ht,max ) , (0.5 lt , −ht,max ) ,
(0.8 lt , −ht,max ) , (0.8 lt , 0) and (lt , 0) , where lt and ht,max denote the length
u1 = uRe on Γ,l Γr . (21)
and the maximum height of the texture, respectively. The coordinates
When (21) is imposed on both Γl and Γr , we enforce zero mean of a point on the B-spline curve are then for t ∈ [0,1] determined by
pressure to obtain a unique pressure solution p1. However, with (21) we
observed an unstable behavior, i.e., oscillations in the solution (u1, p1 ) . ⎧ 16t 3 − 9.6t 2 + 2.4t if 0 ≤ t < 0.25
Consequently, the system (20) has to be stabilized. Instead, we replace x = lt − 3.2t 3 + 4.8t 2 − 1.2t + 0.3 if 0.25 ≤ t < 0.75
⎨ 3 2
the Dirichlet condition on Γr with the natural condition such that (20) is ⎩ 16t − 38.4t + 31.2t − 7.8 if 0.75 ≤ t ≤ 1 (24)
solved with the boundary conditions
and
u1 = uRe on Γl , (22a)
⎧ 48t 3 − 24t 2 if 0 ≤ t < 0.25
1 ⎪ − 16t 3 + 24t 2 − 12t + 1 if 0.25 ≤ t < 0.5
(∇u1) n − p1 n = 0 on Γr . y = ht,max .
Re (22b) ⎨ 16t 3 − 24t 2 + 12t − 3 if 0.5 ≤ t < 0.75
⎪− 48t 3 + 120t 2 − 96t + 24 if 0.75 ≤ t ≤ 1
For three-dimensional computations, u1 = uRe is also prescribed on ⎩ (25)
the sides of the domain. Due to specifying Dirichlet conditions only on
Simulation results for the two test cases are presented in Sect. 3.3.1.
three of the four boundaries in 2D or on five of the six boundaries in 3D,
Even though the examples are unrealistic regarding the bearing dia-
the solution (u1, p1 ) of the cell problem is unique.
meter of about half a millimeter and the rotational speed of 2⋅106 rpm ,
we chose them because simulation times are moderate, making an ex-
3. Effective problem: The only difference regarding the effective
tensive mesh convergence study possible, and results can be better vi-
problem compared with the original upscaling approach concerns
sualized with only few textures. However, in Sect. 3.3.2, we check
the boundary conditions for the lower boundary. We split the
whether the results transfer to a more realistic setting with 1,500 tex-
boundary into Γ3 and Γeff . On Γ3 , we impose the no-slip condition
tures of the same size such that the bearing has a diameter of about 5 cm
(4a), u = 0 . On Γeff , i.e., on top of the textures, we replace (9) with
and runs at 20,000 rpm . For this third test case, we used the relative
u (x ) = u1, i (x ) , x ∈ Γeff . (23) eccentricity erel = 0.3.
We implemented a Navier-Stokes solver and our effective solver
Hence, for each cell problem i, the respective velocity solution u1, i is
with the help of the C++ finite element library deal.II [22]. For the
evaluated pointwise instead of taking the mean value as in (9), (10).
The quality of this new three-step effective approach is investigated
Table 1
in the following section.
Governing equations (gov. eqs.), boundary conditions (b.c.) and additional
boundary conditions in the three-dimensional case on the sides (3D b.c.) to
3. Numerical results obtain Navier-Stokes, Stokes, Reynolds or effective solutions.
gov. eqs. b.c. 3D b.c.
To demonstrate the potential of our effective Navier-Stokes model,
we investigate two- and three-dimensional test cases for textured Navier-Stokes (2) (3), (4) (3)
journal bearings. We compare simulation results obtained from solving
the Navier-Stokes, Stokes, Reynolds and effective problem. To give an Stokes (2) without (3), (4) (3)
(u⋅∇) u
overview, Table 1 summarizes the particular governing equations and
boundary conditions. Note that the effective model is split into its three Reynolds (11) p periodic p=0
parts Reynolds problem, cell problems and effective problem. Hence,
with effective solution or solution of the effective model/approach we refer effective model
to the solution of the effective problem (2) with boundary conditions
Reynolds (11) p periodic p=0
(3), (4) and (23), requiring the previous solution of the Reynolds and
cell problems (20) (4), (22) u1 = uRe
the cell problems. effective prob. (2) (3), (4), (23) (3)
Due to large rotational velocities, large texture heights and small

251
M. Rom, S. Müller Tribology International 124 (2018) 247–258

Table 2 velocity (18) which has to be a sufficiently accurate approximation of

Parameters for the two-dimensional simulations. the Navier-Stokes velocity, in particular the dominant component u in
journal bearing flow direction x. Fig. 5 shows a comparison of u in the first half of the
diameter d {1.5/ π ; 150/ π } mm unfolded system at the position y = 0.5 c (1 − erel ) = 3.5 μm , i.e., at half
number of textures nt {15; 1,500} the minimum gap height, for the test case nt = 15, erel = 0.3. The
texture height ht,max 30 μm agreement of the curves between the textures is good. This is important
texture length lt 50 μm because the Reynolds velocities at the leading edge of each texture are
distance betw. textures dt , x 50 μm
used to compute the boundary condition (22a) which induces the flow
lubricant
density ρ
field in the cell problem domain. In contrast, the agreement of the
880 kg/m3
dynamic viscosity η 0.01 Pa s
curves over the textures is poor. However, these values for the Reynolds
operating conditions velocity only enter the convective term in the cell problem system (20)
radial clearance c 10 μm and, hence, have less influence. Alternatively, we solved the Stokes
relative clearance ψ=
2c {0.042; 0.00042} system without the convective term instead of (20). However, this led
d
relative eccentricity erel {0.3; 0.8} to worse effective simulation results in all cases for which we tested this
rotational velocity urot 50 m/s option.
revolutions R {2⋅106; 2⋅10 4}/min For comparison, we also considered the upscaling approach and
Reynolds number Re 44 solved the zeroth-order problem in the domain Ω0 , cf. Fig. 2. In Ω0 , the
film thickness is significantly larger than the actual film thickness of the
test case. For instance at x = 0 , the actual film thickness is 13 μm . This
Navier-Stokes as well as for the cell and effective problems, we use
is increased to 43 μm in Ω0 because the maximum texture height of
quadrilateral Q2/ Q1 Taylor-Hood elements (biquadratic velocity/bi-
30 μm is added. This leads to a fundamentally changed velocity dis-
linear pressure elements). For the Reynolds problem, Q2 elements (bi-
tribution regarding the film thickness direction y. Consequently, with
quadratic pressure elements) are chosen. These guarantee a continuous
values between 36.9 and 46.2 m/s at y = 3.5 μm for x between 0 and
pressure gradient which is used to compute the Reynolds velocity, see
0.8 mm , the zeroth-order velocity component u 0 is much larger than the
(14) to (18). The nonlinear Navier-Stokes and effective problems are
Navier-Stokes and Reynolds velocities in Fig. 5.
solved iteratively by applying Picard linearization. The resulting linear
systems and also the linear cell problem systems are solved by applying
3.3. 2D simulation results
the GMRES solver of deal.II. For preconditioning of these systems, we
implemented the BFBt preconditioner, see for instance [23]. The Picard
In the following, we compare Navier-Stokes simulation results with
iteration terminates when the maximum absolute difference between
results of our effective approach, in Sect. 3.3.1 for the two test cases
two consecutive solutions is smaller than a prescribed threshold. For all
erel = 0.3 and erel = 0.8, both with nt = 15. The more realistic test case
simulations described in this paper, threshold values of 0.1 m/s for the
with 1,500 textures and erel = 0.3 is investigated in Sect. 3.3.2.
velocity components and 1,500 Pa for the pressure turned out to be
sufficient. The linear Reynolds problem is solved by the CG solver with
3.3.1. Test cases with 15 textures
SparseILU preconditioner (incomplete LU decomposition for sparse
Fig. 6 shows contour plots of the velocity components u and v in the
matrices) of deal.II. The simulations were performed on a machine with
area of the seventh texture for erel = 0.8. As in Fig. 4, the leading and the
32 Intel Xeon E5-2667 v3 @ 3.2 GHz CPUs and 256 GB RAM. The im-
trailing edge of the texture are indicated. The Navier-Stokes solutions
plementations are serial, but some parts of the deal.II solvers are
on the top can be directly compared with the cell problem solutions in
thread-parallel.
the middle. Differences in the flow field are clearly visible, but the
At first, we studied mesh convergence for the Navier-Stokes and
general flow behavior with the recirculation in the texture is similar.
effective simulations, i.e., we conducted simulations with uniformly
The only data of the cell problem used in the effective problem are the
increasing mesh resolution until the differences between the solutions
velocity values on the line y = 0 which enter the effective problem as
of two consecutive mesh refinement levels were negligible. A part of the
final unfolded Navier-Stokes mesh around the fourth texture for the low
eccentricity test case nt = 15, erel = 0.3 is depicted on the top of Fig. 4,
which also illustrates the B-spline texture shape. The thick vertical lines
indicate the leading and the trailing edge of the texture. In these areas,
the mesh resolution is higher to resolve the occurring pressure peaks.
For the effective approach, we have to distinguish between three dis-
cretizations. (i) The Reynolds equation is discretized very fine since the
computational effort for this one-dimensional problem is negligible. (ii)
The necessary mesh resolution for the cell problems turned out to be the
same as for the corresponding parts in the Navier-Stokes mesh. In Fig. 4
(top), this is the texture part in between the thick vertical lines. (iii) The
effective mesh can be chosen coarser than the Navier-Stokes mesh re-
garding the film thickness direction y to obtain mesh-converged results.
The number of mesh cells in flow direction x is the same for both me-
shes, but the cells of the effective mesh are distributed more uniformly
over the texture. Overall, the effective mesh needs less than half the
number of cells and degrees of freedom of the Navier-Stokes mesh. A
part of the final effective mesh is shown on the bottom of Fig. 4. The
final numbers of mesh cells and degrees of freedom are listed in Table 3.

3.2. Approximation quality of the Reynolds velocity

Fig. 4. Part of the Navier-Stokes mesh (top) and corresponding part of the ef-
A crucial ingredient of our effective approach is the Reynolds fective mesh (bottom) for nt = 15, erel = 0.3 .

252
M. Rom, S. Müller Tribology International 124 (2018) 247–258

Table 3
Numbers of mesh cells and degrees of freedom for mesh-converged 2D simu-
lation results for nt = 15, erel = {0.3; 0.8} .
mesh cells degrees of freedom

Navier-Stokes 12,480 117,183

Reynolds 1,800 3,601

cell problem 640 6,063
effective problem 5,760 55,483

Fig. 5. Navier-Stokes and Reynolds velocity component u at

y = 0.5 c (1 − erel ) = 3.5 μm for nt = 15, erel = 0.3 .

boundary values, see (23). On this line, the velocities u and v of the
Navier-Stokes and the cell problem solution match sufficiently. Conse-
quently, the effective velocity solutions on the bottom of Fig. 6(a) and
(b) are in good agreement with the corresponding part in the respective
Navier-Stokes solution on the top.
Figs. 7 and 8 for erel = 0.3 and erel = 0.8, respectively, support the
aforementioned findings. They again illustrate the velocity components
u and v but now for the first eight textures along the line y = 0 , i.e., the
figures show the boundary condition (23). Note the different scales on
the y-axes of the plots. Due to the different eccentricities, the two test
cases strongly differ regarding the distribution and the magnitudes of
the velocities. A higher eccentricity leads to a smaller minimum film
thickness and a higher acceleration in the convergent part. Hence, the
velocity component u in flow direction x is larger. In addition, the
squeeze effect and the recirculation in the textures are stronger such
that the velocity component v in film thickness direction y is also larger.
The general behavior of the four Navier-Stokes velocity curves is
matched adequately by the effective model. Considerable deviations
only concern the velocity peaks. The maximum percentage deviation in
the velocity component u for erel = 0.3, see Fig. 7 (top), is about 10 %.
For erel = 0.8, see Fig. 8 (top), there is one outlier: at the fifth texture,
the deviation between the positive velocity peaks is 61 %. Apart from
this outlier, the maximum deviation is about 23 %. Overall, the devia-
tions seem to increase with increasing eccentricity. This can probably
be explained by the Reynolds velocity entering the cell problems which
as an approximation of the Navier-Stokes velocity becomes worse with
increasing velocity magnitudes.
For the velocity component v, the picture is not so clear. Here, the
test case erel = 0.3, see Fig. 7 (bottom), has the higher maximum relative
deviation, namely 55 % at the eighth texture. For erel = 0.8, see Fig. 8
(bottom), the maximum is 28 % at the second texture. However, the
absolute deviations are larger for the high eccentricity case which fits
the expectation since the Reynolds velocity becomes a worse approx-
imation as explained above.
Regarding the quality of the pressure distribution, being
(caption on next page)

253
M. Rom, S. Müller
Fig. 6. Velocity components (a) u and (b) v in the area of the seventh texture for
nt = 15, erel = 0.8 : Navier-Stokes (top), cell problem (middle) and effective so-
lution (bottom).
Fig. 7. Velocity components u (top) and v (bottom) for the first eight textures
along the line y = 0 for nt = 15, erel = 0.3 .
Fig. 8. Velocity components u (top) and v (bottom) for the first eight textures
along the line y = 0 for nt = 15, erel = 0.8 .
investigated next, it seems to be important that the general behavior of
the velocities is met adequately by the effective model, whereas the
exact reproduction of the velocity peaks is of minor relevance.
Comparisons of the pressure solutions along y = 0 are shown in
Figs. 9 and 10 for erel = 0.3 and erel = 0.8, respectively, together with the
film thickness h as an orientation. These plots additionally contain the
Reynolds and the Stokes solution and clearly demonstrate that the
Reynolds equation and the Stokes system are not valid for our test
problems with deep textures of moderate length and large rotational
velocity. In contrast, the effective pressure fits the Navier-Stokes pres-
sure well. Only the pressure distribution over the central texture, i.e.,
around x = 0.5 πd = 0.75 mm , in the case erel = 0.8 shows a consider-
able deviation.
254
Tribology International 124 (2018) 247–258
In the following, the good qualitative results from Figs. 6–10 are
substantiated by a quantitative evaluation. Therefore, we compare
characteristic values, namely the load-carrying capacity
W: = ∫ [p]+ dγ ,
Γ4 (26)
where we only consider positive values [p]+ = max(0, p) of the pressure
p, the friction force
F: = ∫ τ dγ
Γ4 (27)
with the shear stress τ and the volumetric flow rate
Q: = ∫ u dγ ,
Γ2 (28)
which are here given in dimensioned form. Due to the two-dimensional
setting, the characteristic values are all determined per unit width. Note
that in the case of the Reynolds problem, (26) to (28) are replaced by
W : =∫Ω [p]+ dx , F : =∫Ω τ dx and Q: =∫Γ uRe dγ , respectively.
Re Re per
Table 4 lists the results for the Navier-Stokes simulations for both test
cases and the percentage deviations from these values for the effective,
Reynolds and Stokes solution.
For the first test case erel = 0.3, the effective and the Stokes solution
show similar relative deviations of less than 1 % from the Navier-Stokes
solution. Hence, the Stokes system could be considered valid in this
case. However, a closer look at the pressure distribution in Fig. 9 re-
veals that the Navier-Stokes pressure solution is slightly asymmetric,
which is induced by inertia effects as observed in, e.g., [24,25]. Since
this asymmetry cannot be reproduced by the Stokes solution, the
agreement of the Navier-Stokes and Stokes pressures is only good in the
first half of the bearing where the pressure is positive. In contrast, the
effective solution can reproduce the asymmetry. The Reynolds equation
is not valid as can be deduced from the plot and from the deviation of
the load-carrying capacity W of 8.67 % .
For the high eccentricity test case erel = 0.8, all deviations from the
Navier-Stokes solution increase. With 17.12 % and 7.99 % regarding the
load-carrying capacity W, both Reynolds and Stokes solution cannot be
considered to be reasonable approximations of the Navier-Stokes so-
lution. In contrast, the effective solution only differs by about 1 % .
3.3.2. Test case with 1,500 textures
As described in Sect. 3.1, it remains to demonstrate the applicability
to a more realistic setting regarding the bearing size and the rotational
speed. In the following, we investigate the test case erel = 0.3 with 1,500
instead of 15 textures of the same geometry and with the same distance
between the textures as before. Hence, the diameter of the bearing in-
creases from 1.5/ π mm to 150/ π mm and the rotational speed decreases
from 2⋅106 rpm to 20,000 rpm .
Fig. 9. Pressure p and film thickness h along the line y = 0 for nt = 15, erel = 0.3 .
M. Rom, S. Müller Tribology International 124 (2018) 247–258

pointwise as boundary conditions, an adequate discretization of the

effective problem is also important. Hence, for reasonable effective
results, the construction of suitable meshes is essential.

3.4. 3D simulation setups

The simulation parameters for the three-dimensional test case are

listed in Table 6. Many of them are unchanged compared with the two-
dimensional examples. The bearing surface contains three rows of
textures with twelve textures in each row. For simplicity of mesh gen-
eration, the textures are cuboidal with a constant height of 40 μm .
For the three-dimensional Navier-Stokes simulation, we constructed
a mesh according to the results of the mesh convergence investigation
of Sect. 3.1, i.e., with similar mesh spacings in x- and y-direction. The
spacing in the additional z-direction was chosen slightly finer than the
Fig. 10. Pressure p and film thickness h along the line y = 0 for nt = 15, spacing in x-direction. The computational effort in 3D makes it im-
erel = 0.8 . possible to conduct a full mesh convergence study, which would be
difficult anyway due to the pressure singularities occurring at the sharp
Table 4 texture edges and corners. We adapted the Navier-Stokes mesh to these
Comparison of load-carrying capacity W [N/m], friction force F [N/m] and singularities by using a stretching of the mesh cells toward the edges
volumetric flow rate Q [c m2/s], each given per unit width, for nt = 15, and corners, see Fig. 12.
erel = {0.3; 0.8} . The discretizations of the Reynolds, cell and effective problems were
erel Nav.-St. effective Reynolds Stokes also set up conforming to the main findings of the mesh convergence
study of Sect. 3.1. (i) Again, a very fine discretization was chosen for the
0.3 W 770.53 + 0.73 % − 8.67 % + 0.93 % Reynolds problem since the computational effort for this two-dimen-
F 81.49 + 0.63 % − 5.61 % − 0.70 %
sional problem is negligible compared with the three-dimensional
Q 2.34 − 0.38 % − 1.71 % + 0.10 %
0.8 W 3,309.09 + 1.03 % − 17.12 % − 7.99 % problems. (ii) As in 2D, the cell problem meshes exactly agree to the
F 184.44 + 1.23 % − 10.09 % − 2.09 % corresponding parts in the Navier-Stokes mesh, i.e., they were con-
Q 0.74 − 1.41 % − 3.62 % − 1.41 % structed as if they were cut out of the Navier-Stokes mesh. (iii) Re-
garding the effective mesh, the main result was the need for less re-
solution in film thickness direction y compared with the Navier-Stokes
The pressure along y = 0 is depicted in Fig. 11 together with a close-
up of the area of the pressure maximum. In contrast to the corre-
sponding test case with 15 textures, the pressure distribution is almost
symmetric. While an effect of the textures on the pressure distribution is
not visible in the plot on the top of Fig. 11, it can be clearly seen in the
close-up on the bottom. Again, the effective solution is in good agree-
ment with the Navier-Stokes solution, whereas the Reynolds and the
Stokes solution show significant deviations.
A quantitative comparison of Navier-Stokes, effective, Reynolds and
Stokes solution is given by the comparison of the characteristic values
(26) to (28) in Table 5. The values and percentage deviations for the
friction force F scale from the test case with 15 textures: due to using a
bearing with 100 times the diameter in the case with 1,500 textures, the
friction force is also multiplied by 100. The volumetric flow rate Q stays
the same and also the deviations from it are similar. The load-carrying
capacity W is much larger than in the case with 15 textures. Note that it
is given here in kN/m . While the effective solution provides an even
better approximation of W than in the small test case, the Reynolds and
the Stokes solution exhibit significantly larger deviations of 14.11 % and
5.13 % , respectively. This is due to the almost symmetric pressure dis-
tribution such that the deviations in the positive pressure part are ap-
proximately the same as the corresponding deviations in the negative
part. This was not the case for the small example with 15 textures, cf.
Fig. 9.
The test case with 1,500 textures revealed a sensitivity of the effec-
tive approach regarding the quality of the meshes for the cell and ef-
fective problems. Small adjustments of the mesh points which affect the
mesh cell quality, e.g., the orthogonality of the cells, induce small
modifications of the cell problem solutions. In turn, these modifications
can trigger substantial changes of the pressure distribution computed
from the effective problem. This effect is not apparent in a case with
only a few textures but adds up more and more with an increasing
number of textures. In particular, the mesh cells around the line y = 0 ,
where data are extracted for the effective problem, should be of good Fig. 11. Pressure p along the line y = 0 for nt = 1,500 , erel = 0.3 (top) with close-
quality. Since the cell problem solutions enter the effective problem up at the pressure maximum (bottom).

255
M. Rom, S. Müller Tribology International 124 (2018) 247–258

Table 5 Navier-Stokes solution on the top. This is substantiated by the plot in

Comparison of load-carrying capacity W [kN/m], friction force F [N/m] and Fig. 14 which depicts the pressure along the line
volumetric flow rate Q [cm2/s], each given per unit width, for nt = 1,500 , y = 0, z = 1/2 wb = 0.105 mm . Apart from the pressure peaks, i.e., the
erel = 0.3 . singularities, the effective curve agrees well with the Navier-Stokes
erel Nav.-St. effective Reynolds Stokes curve. In contrast, similarly to the two-dimensional case, the Reynolds
and the Stokes solution deviate noticeably from the Navier-Stokes so-
0.3 W 8,217.7 + 0.16 % − 14.11 % − 5.13 %
lution. This is again due to using deep textures of moderate length and a
F 8,148.6 + 0.63 % − 5.60 % − 0.70 %
Q 2.34 − 0.33 % − 1.73 % + 0.10 %
large rotational velocity such that Reynolds equation and Stokes system
cannot reflect the recirculation and inertia effects. Furthermore, both
cannot reproduce the asymmetry of the pressure distribution induced
Table 6 by inertia.
Parameters for the three-dimensional simulations. An effect which is only present in the three-dimensional setting is
the cross flow resulting in side leakage of the journal bearing. Fig. 15
journal bearing
diameter d 1.5/ π mm
presents a comparison of the velocity component w which confirms that
width wb 0.21 mm the cross flow can be adequately captured by the effective simulation.
number of textures nt 36 (12⋅3) A comparison of the characteristic values (26) to (28) can be found
texture height ht 40 μm in Table 8. Again, all deviations of the effective values from the Navier-
texture length lt 60 μm
Stokes values are below 1 % , whereas the Reynolds and Stokes values, in
texture width wt 30 μm
particular the load-carrying capacity W, indicate the nonapplicability of
distance betw. text. (x-dir.) dt , x 65 μm
distance betw. text. (z-dir.) dt , z 40 μm the latter two solvers.
lubricant
density ρ 880 kg/m3 3.6. Performance of the effective Navier-Stokes model
dynamic viscosity η 0.01 Pa s
operating conditions
For the evaluation of the performance of our effective approach, we
radial clearance c 10 μm
relative clearance 2c 0.042 compare the computation times for the four different test cases de-
ψ=
d scribed in Sects. 3.1 to 3.5. The results are summarized in Table 9 by
relative eccentricity erel 0.5
giving the computation times in seconds of the Navier-Stokes and the
rotational velocity urot 50 m/s
revolutions R 2⋅106/min
effective simulations, where the latter are split into the times for solving
Reynolds number Re 44 Reynolds, cell and effective problems. The last column of the table
contains the overall speed-up of the effective method, i.e., the speed-up
is computed for each simulation by dividing the Navier-Stokes value by
the sum of the Reynolds, cell and effective values.
The best speed-ups are obtained for the two-dimensional test cases
with 15 textures. Due to the small number of textures and, hence, cell
problems, the dominating part of the effective simulation is solving the
effective problem. The high speed-ups of about 8 can be explained by
the coarser mesh for the effective problem which saves approximately
53 % of the degrees of freedom compared with the Navier-Stokes mesh,
cf. Table 3.
The two-dimensional example with 1,500 textures is a special case:
Fig. 12. Part of the Navier-Stokes mesh for the three-dimensional simulation. due to large numbers of degrees of freedom (about twelve million for
the Navier-Stokes simulation) and using codes which in principle are
serial, all linear systems were solved with the direct solver
mesh. However, to be on the safe side, we did not reduce the resolution
SparseDirectUMFPACK of deal.II instead of the iterative GMRES solver
for the 3D example. Hence, as effective mesh the part of the Navier-
to save computation time. The speed-up drops to 2.9 for two reasons: (i)
Stokes mesh without the textures ( y ≥ 0 ) was chosen. Table 7 sum-
due to the large number of textures, more computation time is needed
marizes information on the final discretizations.
for solving the cell problems than for solving the effective problem and
When conducting the effective simulations, we observed that the
(ii) using the direct instead of the iterative solver pays off more for
effective solution is strongly influenced by the choice of the boundary
solving the complex effective problem than for solving the simple cell
conditions for the cell problems. Recall the boundary conditions (22)
problems with much less degrees of freedom. Both reasons lead to the
originally used in the two-dimensional case. For 3D, these were ex-
insight that it is important to keep the time spent for solving the cell
tended by prescribing u1 = uRe not only on Γl but also on the sides of the
problems at a minimum for the overall effective approach to be bene-
domain. However, an improvement of the effective results could be
ficial. Therefore, a reduced basis method could be implemented, simi-
obtained by specifying an alternative strategy: each cell problem was
larly as in Ref. [10] in the context of homogenization of the Reynolds
solved twice, (i) with u1 = uRe on Γl and the sides and the natural con-
equation or in Ref. [26] in the context of the compressible Navier-
dition (22b) on Γr and (ii) with u1 = uRe on Γr and the sides and the
natural condition on Γl . Subsequently, the cell problem solution was set
Table 7
to be the average of these two solutions. For a conclusive statement
Numbers of mesh cells and degrees of freedom for the three-dimensional si-
about the best boundary conditions for the cell problems, further stu- mulations.
dies are necessary. These should also take into account a stabilization of
the flow in the case of Dirichlet conditions on all boundaries. mesh cells degrees of freedom

Navier-Stokes 331,776 8,980,116

3.5. 3D simulation results
Reynolds 88,200 354,295
cell problem 4,096 114,444
Fig. 13 shows the pressure p for an extract of the bearing surface effective problem 221,184 5,939,700
( y = 0 ). The effective solution on the bottom is very similar to the

256
M. Rom, S. Müller Tribology International 124 (2018) 247–258

Table 9
Computation times in seconds of Navier-Stokes and effective simulations
(Reynolds + cell + effective problems) and overall speed-up of effective ap-
proach. The test cases are determined by dimension, number of textures nt and
relative eccentricity erel .
test case Nav.-St. effective approach sp.-up

Re. cell eff.

2D/15 804 0.03 11 85 8.4

2D/15 1,207 0.03 13 126 8.7
2D/1,500/0.3 2,426 0.83 436 388 2.9
3D/36 225,450 26 5,672 41,141 4.8

Fig. 13. Pressure p on an extract of the bearing surface y = 0 : Navier-Stokes
(top) and effective solution (bottom).
the case with 1,500 textures, the computation time reduces from 40 to
14 min. When it comes to a repeated application, e.g., for finding the
optimal texturing for a journal bearing, the reduction adds up and,
hence, becomes appreciable quickly.
In 3D, it is not possible to apply the direct solver because its memory
consumption is too large. Using the iterative solvers led to a speed-up of
4.8 which is smaller than in the 2D cases with 15 textures. This is due to
being cautious in the 3D case and using a probably over-refined effec-
tive mesh as described in Sect. 3.4. In 3D, we only save about 34 % of
degrees of freedom instead of 53 % in the 2D case. However, the speed-
up of 4.8 means a significant reduction of computation time from 62 to
13 h.

4. Conclusion

Our new effective Navier-Stokes model has been designed for the
investigation of textured surface lubrication, in particular for applica-
tions with deep textures of moderate length and high flow velocities.
Then, the Reynolds equation and the Stokes system fail to provide ac-
curate results such that the Navier-Stokes equations have to be con-
Fig. 14. Pressure p and film thickness h along the line sidered instead. The basis for our effective method is the application of
y = 0, z = 1/2 wb = 0.105 mm . an upscaling technique to the incompressible Navier-Stokes system.
However, since the texture height is not necessarily of small scale
compared with the film thickness, a modification of the original up-
scaling approach had to be found to obtain reasonable results. The main
idea was the replacement of the zeroth-order velocity solution with the
Reynolds velocity which provides a sufficient approximation of the
Navier-Stokes velocity, at least between the textures.
All two- and three-dimensional examples of textured journal bear-
ings investigated in this paper have shown that our effective approach
leads to accurate computations while considerably reducing the com-
putational cost compared with solving the Navier-Stokes equations.
There is still potential for improvement of our model. Regarding the
accuracy, the Reynolds velocity computed over the textures could be
improved, e.g., by finding a suitable damping factor. In addition, the
Fig. 15. Velocity component w in the middle of the third texture column boundary conditions for the cell problems have to be investigated fur-
( x = 0.3125 mm ): Navier-Stokes (top) and effective solution (bottom). ther. For the three-dimensional simulations, it is not clear why the al-
ternative strategy of solving each cell problem twice as described in
Sect. 3.4 led to the best effective results. Instead of using a combination
Table 8
of Dirichlet and natural boundary conditions, prescribing Dirichlet
Comparison of load-carrying capacity W [N], friction force F [N] and volu-
metric flow rate Q [cm3/s] for the three-dimensional test case.
conditions for the velocity on all boundaries and stabilizing the cell
problem system should also be taken into account. Regarding the per-
erel Nav.-St. effective Reynolds Stokes formance of our effective method, further improvements could be
0.5 W 0.0328 + 0.26 % − 22.62 % − 8.37 % achieved by parallelizing the implementation and by applying a re-
F 0.0179 + 0.47 % − 6.59 % − 0.93 % duced basis method to the cell problems.
Q 0.0786 − 0.02 % + 0.52 % + 0.32 % For more realistic simulation results, our effective approach would
have to be extended by a cavitation model.

Stokes equations. An additional improvement could be obtained by a

Acknowledgments
full parallelization of the effective approach which would be advanta-
geous in particular for solving the cell problems since these are in-
The authors would like to thank Dr. K. Backhaus and Prof. G. Knoll,
dependent of each other. However, even with a speed-up of only 2.9 in
IST mbH, for fruitful discussions of simulation parameters and results.

257
M. Rom, S. Müller
This research project is funded by the Central Innovation
Programme for SMEs (ZIM) of the German Federal Ministry for
Economic Affairs and Energy (BMWi), grant number: ZF4151601LP5.
References
[1] Etsion I. State of the art in laser surface texturing. J Tribol 2005;127:248–53.
[2] Cupillard S, Glavatskih S, Cervantes MJ. Computational fluid dynamics analysis of a
journal bearing with surface texturing. Proc Inst Mech Eng Part J: J Eng Tribol
2008;222:97–107.
[3] Gadeschi GB, Backhaus K, Knoll G. Numerical analysis of laser-textured piston-rings
in the hydrodynamic lubrication regime. J Tribol 2012;134:041702.
[4] Taee M, Torabi A, Akbarzadeh S, Khonsari MM, Badrossamay M. On the perfor-
mance of EHL contacts with textured surfaces. Tribol Lett 2017;65:85.
[5] Rahmani R, Rahnejat H. Enhanced performance of optimised partially textured load
bearing surfaces. Tribol Int 2018;117:272–82.
[6] Gropper D, Wang L, Harvey TJ. Hydrodynamic lubrication of textured surfaces: a
review of modeling techniques and key findings. Tribol Int 2016;94:509–29.
[7] Reynolds O. On the theory of lubrication and its application to Mr. Beauchamp
Tower's experiments, including an experimental determination of the viscosity of
olive oil. Phil Trans Roy Soc Lond 1886;177:157–234.
[8] Patir N, Cheng HS. An average flow model for determining effects of three-di-
mensional roughness on partial hydrodynamic lubrication. J Lubr Technol
1978;100:12–7.
[9] Almqvist A, Dasht J. The homogenization process of the Reynolds equation de-
scribing compressible liquid flow. Tribol Int 2006;39:994–1002.
[10] Rom M, Müller S. A reduced basis method for the homogenized Reynolds equation
applied to textured surfaces. Commun Comput Phys 2018https://doi.org/10.4208/
cicp.OA-2017-0121.
[11] Sahlin F, Glavatskih SB, Almqvist T, Larsson R. Two-dimensional CFD-analysis of
micro-patterned surfaces in hydrodynamic lubrication. J Tribol 2005;127:96–102.
[12] Feldman Y, Kligerman Y, Etsion I, Haber S. The validity of the Reynolds equation in
modeling hydrostatic effects in gas lubricated textured parallel surfaces. J Tribol
258
Tribology International 124 (2018) 247–258
2006;128:345–50.
[13] Dobrica MB, Fillon M. About the validity of Reynolds equation and inertia effects in
textured sliders of infinite width. Proc Inst Mech Eng Part J: J Eng Tribol
2009;223:69–78.
[14] Zhang H, Hua M, Dong GN, Zhang DY, Chin KS. A mixed lubrication model for
studying tribological behaviors of surface texturing. Tribol Int 2016;93:583–92.
[15] Ma C, Duan Y, Yu B, Sun J, Tu Q. The comprehensive effect of surface texture and
roughness under hydrodynamic and mixed lubrication conditions. Proc Inst Mech
Eng Part J: J Eng Tribol 2017;231:1307–19.
[16] de Kraker A, van Ostayen RAJ, van Beek A, Rixen DJ. A multiscale method mod-
eling surface texture effects. J Tribol 2007;129:221–30.
[17] de Kraker A, van Ostayen RAJ, Rixen DJ. Development of a texture averaged
Reynolds equation. Tribol Int 2010;43:2100–9.
[18] Shyu SH, Hsu WC. A numerical study on the negligibility of cross-film pressure
variation in infinitely wide plane slider bearing, Rayleigh step bearing and micro-
grooved parallel slider bearing. Int J Mech Sci 2018;137:315–23.
[19] Caramia G, Carbone G, De Palma P. Hydrodynamic lubrication of micro-textured
surfaces: two dimensional CFD-analysis. Tribol Int 2015;88:162–9.
[20] Deolmi G, Dahmen W, Müller S. Effective boundary conditions: a general strategy
and application to compressible flows over rough boundaries. Commun Comput
Phys 2017;21:358–400.
[21] Hamrock BJ, Schmid SR, Jacobson BO. Fundamentals of fluid film lubrication.
second ed. Boca Raton: CRC Press; 2004.
[22] Bangerth W, Hartmann R, Kanschat G. deal.II - a general purpose object-oriented
finite element library. ACM Trans Math Software 2007;33(24):1–27.
[23] Elman HC. Preconditioning strategies for models of incompressible flow. J Sci
Comput 2005;25:347–66.
[24] Yamada Y, Nakabayashi K. On the flow between eccentric rotating cylinders when
the outer cylinder rotates. Bull JSME 1968;11:455–62.
[25] Arghir M, Roucou N, Helene M, Frene J. Theoretical analysis of the incompressible
laminar flow in a macro-roughness cell. J Tribol 2003;125:309–18.
[26] Deolmi G, Müller S. A two-step model order reduction method to simulate a com-
pressible flow over an extended rough surface. Math Comput Simulat
2018;150:49–65.