Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: A parallel algorithm is presented for the Direct Numerical Simulation of convection flows in open or
Received 31 December 2017 partially confined periodic domains, containing immersed cylindrical bodies of arbitrary cross-section.
Revised 22 January 2019
The governing equations are discretized by means of the Finite Volume method on Cartesian grids.
Accepted 1 March 2019
The method presented includes a triperiodic Poisson solver employed irrespective of the actual
Available online 2 March 2019
boundary shape and a second order accuracy for the computational domain, including the near wall re-
Keywords: gions, when walls are defined as immersed boundaries.
Direct Numerical Simulation The numerical solution of the set of linear equations resulting from discretization is carried out by
Free convection means of efficient and highly parallel direct solvers. Verification and validation of the numerical proce-
Immersed Boundary Method dure is reported in the paper, for laminar and turbulent pipe flow, and for the case of flow around an
array of heated cylindrical rods arranged in a triangular lattice. The formal accuracy of the method is
demonstrated in laminar flow conditions, and DNS results in turbulent conditions are compared to avail-
able literature data, thus confirming the favorable qualities of the method.
© 2019 Elsevier Ltd. All rights reserved.
1. Introduction ulate flow over bodies that, in general, do not conform to the
computational grid. These include methods such as the immersed
While the continuous increase in available computing resources interface method [4], the Cartesian grid method [5] and the
as well as novel mathematical, algorithmic and coding techniques cut-cell method [6]. Most of these variants are listed and briefly
has allowed DNSs to be performed at very high Reynolds numbers described in the review by Mittal and Iaccarino [7]. The main
[1] or with extremely refined grids [2], DNSs are often limited to drawback of many of these approaches is in the accurate represen-
flows in very simple domains like parallel channels, pipes or flat tation of boundary conditions close to irregular boundaries, where
plates, mainly due to the limitations of some of the most widely the increase in local truncation error can affect the global accuracy
used discretization methods for the DNS (spectral methods) and to of spatial discretization.
the numerical errors that often affect the resolution of the discrete In more recent years the problem of the accuracy of spatial
equations on unstructured grids. The search for parallel and accu- schemes and of the representation of the boundaries has been con-
rate methods for the DNS of flows in complex geometries is still sidered and in good part resolved for finite difference [8–11] and
a relevant topic of investigation in the field of turbulent flow and finite element [12,13] methods, also at high order of accuracy, but
heat transfer. to the authors’ knowledge it is difficult to find immersed boundary
Approaches based on the representation of arbitrary domains methods which, in a Finite Volume context are capable of preserv-
in the frame of a Finite Difference or Finite Volume Cartesian ing the second order accuracy also close to the boundaries.
grid, commonly known as Immersed Boundary Methods, are good In this work, a novel methodology is presented for the Di-
candidates given the easy mesh generation procedure and the rect Numerical Simulation of fully-developed, buoyancy-induced
straightforward Cartesian parallelization. Stemming from the origi- flow and heat transfer in open or partially confined periodic do-
nal development of the Immersed Boundary Methodby Peskin [3], mains containing prismatic solid bodies of arbitrary cross-section
a variety of different methodologies have been developed to sim- but aligned with one grid direction. Fluid domains can therefore be
built from single or many non overlapping two dimensional shapes
by orthogonal extrusion. Examples of suitable domain shapes are:
∗
Corresponding author. fully developed or axially homogeneous straight channels of ar-
E-mail address: enrico.stalio@unimore.it (E. Stalio). bitrary cross section (pipes, ducts of triangular cross section,...)
https://doi.org/10.1016/j.compfluid.2019.03.002
0045-7930/© 2019 Elsevier Ltd. All rights reserved.
D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159 149
including multiply connected sections (flow about tube bundles 2.2. Momentum and energy equations
where pipe sections are arbitrary) and, by application of wall
boundary conditions on all boundaries, three-dimensional closed A formulation of momentum equations, excluding sources, is
cavities containing internal heaters.
∂ ui ∂ ui ∂p ∂ 2 ui
The continuity, Navier-Stokes and energy equations are dis- ρo + uj =− +μ (2)
cretized by means of the Finite Volume method on Cartesian grids. ∂t ∂xj ∂ xi ∂x j∂x j
Grids are uniform along the homogeneous direction x1 and can
Temperature is solved for in the energy equation
be non-uniform along x2 and x3 . The representation of arbitrar-
ily irregular cylindrical boundaries on Cartesian grids is an ex- ∂T ∂T ∂ 2T
ρ oc p + ui =λ + σ, (3)
tension of the method developed by Barozzi et al. [14] for pure ∂t ∂ xi ∂ xi ∂ xi
diffusion problems. Setting boundary conditions on irregular solid
boundaries involves the definition of a 5-point computational sten- where σ represents a generic source term, expressed in W/m3 .
cil where boundary segments intersect the stencil arms. The dis- For statistically steady flows, thermally and hydrodynamically
cretization method locally uses Finite Differences in order to accu- fully developed in the x1 direction, governing Eqs. (1)–(3) can be
rately enforce conservation laws in the Finite Volume framework arranged to allow for the representation of all variables, including
of this approach. The variables on the modified stencil are mapped temperature and pressure through streamwise periodic quantities,
on the global grid, by means of a linear operator determined by see for example Ref. [15] where also buoyancy effects are consid-
geometrical features and boundary conditions. The overall accuracy ered. This is in view of the computational efficiency of the numer-
of the method is virtually preserved, as well as the computational ical code.
efficiency of the Cartesian approach. For the statistically steady case under scrutiny which is fully
Validation is done on three different test cases: the formal ac- developed along direction x1 , a time and superficial average is de-
curacy of the method is demonstrated against the analytical solu- fined for the generic scalar field f and indicated by angular brackets
tion of a laminar pipe flow. Accuracy at the boundary is further as-
tb
sessed by comparing laminar results of a triangular array of heated 1
< f > ( x1 ) ≡ f (x, t ) d dt (4)
rods against previous studies. Finally, a Direct Numerical Simula- (tb − ta ) ta
tion is performed for the case of a turbulent pipe flow at Reτ =
180, confirming the favorable qualities of the present method and where is the surface of a cross-flow plane (perpendicular to x1 ,
its adequacy for turbulent, fully-developed flow simulation in non- the direction of full development) while [ta , tb ] is a sufficiently
rectangular domains. large time interval for < f > to be independent of time. By aver-
A computational fluid dynamics method is presented includ- aging Eq. (2) for i = 1 and taking the x1 derivative of the result it
ing an Immersed Boundary algorithm which is meant for a Finite is found that the time average of the pressure field must be linear
Volume framework and whose second order accuracy is demon- in x1
strated up to the walls. This algorithm does not implement any d< p >
iterative procedure for adjusting a forcing term, instead it is based =γ (5)
dx1
upon the solution to small linear system of equations at the im-
mersed boundaries, see Section 3.2.1. Its parallel implementation for some constant quantity γ which is negative for flows in the
is based upon a domain decomposition method along the homoge- positive x1 direction. A linear function p(x1 ) can be defined with
neous direction x1 and is very efficient, see Appendix A. The idea respect to a fixed point x01 along the x1 -axis
is to decouple the solution to the Poisson problem along x1 and
p(x1 ) = < p >(x01 ) + γ
x1 − x01 (6)
to adopt a solution strategy for the viscous and the thermal dif-
fusion terms which are implicit only in the orthogonal plane. As which allows to split the pressure into a linear part
p, and a fluc-
this is the plane where the mesh is often stretched for appropri- tuation p :
ately solving boundary layers, stability limits imposed on timestep
p(x1 ) + p (x, t ),
p( x , t ) = (7)
by the discrete form of the diffusion term are typically not very
restrictive. where < p >(x1 ) = 0.
The form of the momentum equations to be solved is then
2. Mathematical formulation
∂ ui ∂ ui ∂ p ∂ 2 ui
ρo + uj =− +μ − γ δi1 (8)
The methodology implements a discrete form of the in- ∂t ∂xj ∂ xi ∂x j∂x j
compressible Navier-Stokes and energy equations for hydro-
In the form of Eq. (8), periodic boundary conditions may be en-
dynamically and thermally fully-developed flows. All the fluid
forced along the streamwise direction on all variables, including
properties are assumed constant, the rate of heating due to vis-
the fluctuating pressure p .
cous dissipation is neglected. Fully developed flow along the x1 -
The momentum equation is written in terms of dimensionless
direction is specifically considered for the formulation of govern-
quantities as
ing equations, but the method presented here does not set specific
requirements for boundary conditions in directions orthogonal to ∂ u∗i ∂ u∗ ∂ p∗ 1 ∂ 2 u∗i ω∗
+ u∗j ∗i = − ∗ + + ∗ δi1 , (9)
the mean flow. ∂t ∗ ∂xj ∂ xi Reτ ∂ x j ∂ x j
∗ ∗
2.1. Continuity equation where the last term equals the wet perimeter divided by the cross
sectional area and it represents the ensemble average of the pres-
Under the above hypotheses, the continuity equation reads: sure drop along the streamwise direction. In Eq. (9), the friction
Reynolds number Reτ is defined as:
∂ ui
=0 (1)
∂ xi Reτ =
uτ Lref
, (10)
ν
where ui indicates the ith component of the velocity field and the
rectangular shape, and metrics of the cells falling near the bound-
ary are conveniently rescaled according to the relative position
∂ϕ
of cell center and boundary, see Fig. 3(a) and (b), therefore the d
¯ + (uϕ ) · n dᾱ = (∇ϕ · n ) dᾱ + q d
¯
union of all Control Volumes of type ¯ which are contained in
¯ ∂t ᾱ ᾱ
¯
the physical domain do not cover the whole domain volume, see (22)
Fig. 2(b).
in case ϕ represents a constant density times the jth component
Utilization of a (for example) lexicographic ordering of the set
of the velocity field, then Eq. (22) forms the momentum balance
of Cartesian Control Volumes defined about stencils of type g will
equation in the jth direction and becomes non-linear.
lead to the development of an efficient algorithm, easily amenable
The advective and diffusive terms of the generic transport
to a parallel implementation. As stencils g do not carry information
Eq. (22) in two dimensions can be approximated as follows
on the position of the domain boundary, it would be impossible in
this case to set the correct boundary conditions. The methodology
4
ᾱ j
presented introduces a linear mapping between stencils of type g (∇ϕ · n ) dᾱ ϕˆ j − ϕˆ 0 ≡ s · ϕˆ (23)
ᾱ δx j
and b so that the discrete form of equations on Control Volumes j=1
¯ is in terms of quantities defined on the grid G and stencils g. In
4
this way, the computational advantages of Cartesian grids is pre- (uϕ ) · n dᾱ ŭ j ϕ̆ j ᾱ j ≡ h · ϕ̆ (24)
served and boundary conditions are set correctly also on irregular ᾱ j=1
boundaries.
where vector ϕ ˆ ≡ (ϕˆ 0 . . . ϕˆ 4 ) is the second-order approximation to
the volume average of ϕ over ¯ and the four adjacent Control Vol-
3.2. Discrete form of the equations on Control Volumes umes while vector ϕ̆ ≡ (ϕ̆1 . . . ϕ̆4 ) contains the second-order ap-
proximation to the superficial average of ϕ over sides of ¯ . Vol-
The transport equation for the continuous scalar function ϕ in- ume and surface averages over and α built around stencils of
tegrated on the generic Control Volume ¯ of external surface ᾱ is type g will be indicated by ϕ and ϕ ˜.
152 D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159
In Eqs. (23) and (24) vector s ≡ (s0 . . . s4 ) contains metric coeffi- With reference to Fig. 4(a) and (b), the index k identifies the
cients. Vector h ≡ (h1 . . . h4 ) contains flow rates across Control Vol- node on the opposite side of k with respect to the center of the
ume boundaries at a specific time. Consistently with the second- stencil, node 0.
order spatial approximation of the method, the midpoint rule for The derivative of function ϕ in k along direction ηk can be ap-
surfaces is used in the advective and diffusive terms, and the mid- proximated using nodes 0, k, k as follows. Taylor-expansion of ϕ
point rule for volumes is used in the unsteady and the source along η and about the point j where j = 0, k, k is
term. Linear differentiation and interpolation are used to recover
∂ϕ 1 ∂ 2ϕ
the algebraic counterparts of the diffusive and advective terms. The ϕ j ϕk + ηj + η2j j = 0, k, k (31)
semi-discrete equation ∂η k
2 ∂η2 k
Fig. 4. Local discretization stencils for intersections with k = 2 (a) and k = 3 (b).
For stencils having a single, different k or multiple intersection The determination of the normal velocity at the k-face, ŭk , is
nodes A, D, c change accordingly. very similar but requires an additional interpolation, because of
It is now possible to formulate the expressions for E and g in the staggering of variables (see Section 3.2.4).
(28):
E = A−1 D, g = A−1 c (41) 3.2.3. Volumetric flow rate across deformed Control Volumes
Depending on the relative distance between the centroid of the
As already mentioned, coefficients in E and g depend on the do-
Control Volume and the intersection point k, three possible con-
main shape and boundary conditions (30) and are calculated once
figurations need to be considered, see Fig. 5. In case (a), the node
at the start of the numerical procedure.
containing the required value of velocity lies within the domain.
The Finite Difference approach close to the boundary is paired
Although this node is no more located on the Control Volume
to the Finite Volume method of the domain interior. It is clear
boundary, due to the deformation induced by the presence of the
from the discussion above that resorting to the finite difference
boundary itself, it can nevertheless be used to determine the value
allows for accurately setting boundary conditions when the do-
of the normal velocity component ŭk . Such an estimation can be
main boundary is not parallel to the Cartesian mesh grid-lines. In
performed by linear extrapolation from the value of the node itself
Section 4.1 this approach is demonstrated to preserve the second
and the value at the intersection point k (which equals zero at a
order accuracy of the method over the whole flow domain for both
solid boundary).
uniform and non uniform mesh. The second order accuracy of the
In case (b), the node containing the required velocity value lies
whole method is ensured by the midpoint rule for surfaces and
outside the domain and is therefore excluded from the calcula-
volumes. Implementation of similar methods at higher orders of
tions. In this situation, a linear interpolation must be performed
spatial accuracy would require a scrupulous distinction between
between the value at the symmetric node of the velocity grid, with
point-wise and space averaged values and the use of specific oper-
respect to the centroid, and the value at the boundary. Such a pro-
ations of integration and deconvolution, see Ref. [19]
cedure is valid, however, only when at least one of the two nodes
An issue remains to be addressed for the implementation of
associated with the stencil direction is contained within the do-
the discrete Eq. (29), this is the determination of vectors h and
main of integration.
ϕ̆, which form a second-order approximation to advective fluxes.
Otherwise, a situation similar to that represented by case (c)
arises: the normal flux on the Control Volume side can not be de-
3.2.2. Advective fluxes termined from the nodal values of the velocity grid, since no “ac-
Implementation of advective fluxes require that Eq. (24) which tive” node falls in the vicinity of the Control Volume itself. The
is reported also here for clarity calculation of the normal velocity at the interface is then operated
4 from the only information available, i.e. the values at the intersec-
(uϕ ) · n dᾱ ŭ j ϕ̆ j ᾱ j tion nodes. In the case of a closed domain, this corresponds to im-
ᾱ j=1 pose zero velocity on the Control Volume side.
is represented numerically within a second order approximation
on both Cartesian and distorted stencils (type-b stencils). 3.2.4. Specific algorithms for the Navier-Stokes equations
The generic surface average ϕ̆ j is expressed as a function of the Staggering. The numerical technique presented here makes use of
volume averaged ϕ value over the central Control Volume ϕ 0 , and staggered grids. In three dimensions this involves four Cartesian
the value of the jth neighbor ϕ j by the following second-order ap- lattices where the union of Control Volumes bounded by grid-
proximation lines of the “main grid” perfectly covers the computational domain,
x j ϕ 0 + x 0 ϕ j aside from immersed boundaries. The three “staggered grids” are
ϕ̆ j = (42) instead displaced by half spacing along x, y and z, respectively.
x j + x 0
Over each of the four grids, volume averaged values and surface
When the stencil is crossed by the domain boundary and j = k, averages over the six faces of each Control Volume are defined for
since the ratio between length of the kth stencil arm and distance all variables. In a numerical code whose spatial accuracy is sec-
of the centroid from the kth Control Volume side is invariant with ond order, the volume averaged values can be identified also with
respect to the stencil deformation, the interpolation weights of dis- point-wise values defined in the volume center, and the same is
torted and Cartesian stencils are the same (see Fig. 3(b)). valid for surface averages.
154 D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159
Fig. 5. Procedure for the estimate of the normal velocity component near the boundary for the energy equation.
If by one side the use of a staggered mesh makes implementa- rithms selected for the fast, parallel solution to the Poisson equa-
tion and bookkeeping more complex, and it involves many inter- tion, are shortly presented in the Appendix A.2.
polations, on the other side the use of different grids for the so-
lution of the energy equation and of each component of momen-
4. Validation and computational performance of the numerical
tum allows for a straightforward evaluation of the momentum and
procedure
energy fluxes across Control Volume faces. But most importantly,
it avoids the occurrence of the so-called “checker-boarding” effect
4.1. Laminar pipe flow
[20] on the variable distributions, which typically arises when a
single mesh is used.
A challenging task for Immersed Boundary Methods on a Carte-
sian mesh is to represent curved boundaries at the prescribed for-
Pressure-Velocity decoupling. It is well known that in the imple- mal order of accuracy. In this regard, the laminar pipe flow is a
mentation of numerical algorithms for incompressible flows, the convenient test case, as it allows for a direct comparison against
continuity equation can be interpreted as a constraint for the com- the analytical solution.
ponents of the velocity vector and procedures are employed for de- A series of tests are carried out at Reτ = 1. In laminar condi-
coupling the calculation of the pressure field from velocity. In the tions the exact solution for the dimensionless streamwise velocity
numerical code presented here, a Projection Method, introduced by component reads as follows:
Chorin [21] and later developed by Gresho et al. [22], is imple-
Reτ
mented. u (r ) = (1 − r 2 ) (43)
Like all fractional step methods, also Projection Methods involve 2
the numerical solution to a Poisson equation per time step, this Both uniform and non-uniform grids are tested, the non-uniform
implies an additional computational cost. The details of the algo- grids have the following hyperbolic tangent distribution:
D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159 155
i−1 1
tanh δ nx − 2 is used, both the centerline velocity and L∞ error decrease with
xi = R δ , (44) x2 , demonstrating that the present technique ensures second-
tanh 2
order spatial accuracy, provided that the mesh is well-suited for
where δ = 2 in the present case. the domain shape, i.e. free of local singularities which occur when
Accuracy results are reported in Fig. 6. For the uniform mesh, the intersections between the boundary and grid-lines lie too close
the centerline velocity approaches the analytical value display- to the cell centers. For simple cases like the turbulent pipe flow
ing second-order accuracy. Local error maxima are independent of described in the following, validation in laminar regime is also a
the mesh size, but depend only on the relative position of the quick, reliable test to check the grid quality before running a full
boundary and grid lines (see Fig. 7). When a nonuniform mesh DNS.
Fig. 8. Triangular rod bundles in laminar conditions: maps of streamwise velocity ux for different P/D ratios and for Reτ = 1, Pr = 1.
156 D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159
ui = 0, i = 1, . . . , 3 (45)
and
∂θ
= −1 (46)
∂η
at the solid walls.
Fig. 8 reports maps of the streamwise velocity component,
while Fig. 9 compares friction factor and Nusselt number values
against correlations reported by Rehme [23]. An excellent agree-
ment is observed between the integral values calculated and exist-
ing correlations.
Fig. 10. Pipe flow validation results for Reτ = 180 and Pr = 0.71.
D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159 157
Fig. 11. Pipe flow DNS at Reτ = 180: instantaneous contours of the streamwise velocity component on a surface at y+ = 10, and isosurfaces of Q.
constant dimensionless time step of t = 10−4 is used. Such pa- the test, including 643 and 2563 Control Volumes, respectively. For
rameters comply with current practice in DNS. the mesh size and number of cores considered, the speedup is sub-
Fig. 10 shows flow and thermal field statistics computed on 100 linear but monotonically increasing.
realizations collected over 50 dimensionless time units. The agree- The computational performance of the procedure is also as-
ment with reference data is deemed as very satisfactory. In order sessed in comparison to a conventional unstructured Finite Vol-
to provide further evidence of the capability of the method to per- ume solver. The general purpose, open-source OpenFOAM toolbox
form resolved DNS of turbulent flow along curved boundaries, in- [27] is selected for the comparison. OpenFOAM is a second-order
stantaneous contours of the streamwise velocity component on a implementation of the PISO algorithm [28] on arbitrarily-shaped
surface at a distance y+ = 10 from the wall are displayed in Fig. 11, polyhedral meshes. One quarter of the cross-section of the 2563
together with isosurfaces of the Q-criterion (as defined by Hunt Cartesian grid on which the present method is applied is shown
et al. [26]). Typical features of turbulent wall bounded flows such in Fig. 13(a), along with a similar visualization of the unstructured,
as low- and high-speed streaks can be observed, along with a mul- hexahedral grid used with OpenFOAM, see Fig. 13(b). Both grids
tiplicity of near-wall structures. were prepared in order to respect typical sizing requirements for
turbulent pipe DNS at Reτ = 180, a summary of the main dimen-
4.4. Scalability and computational efficiency sional features of the two grids is reported in Table 1.
The performance of the two methods is measured in terms of
The computational method presented here is implemented in the cpu time per DoF for the execution of 10 0 0 time steps on
parallel through a domain decomposition technique along the ho- the above-mentioned HPC infrastructure. The number of degrees
mogeneous direction x1 . The parallel implementation is detailed in of freedom (DoF) of a DNS is defined as the number of grid points
the Appendix A. Fig. 12 reports results of a scalability test per- times the number of time steps performed. In order to reason-
formed on a small 4-node, 64-core cluster (Intel Xeon®E5-2680v2 ably compare the performances of the present, fully direct method
2.8Ghz) equipped with an Infiniband network switch, operated by with the OpenFOAM solver, which is iterative, residual tolerances
the High Performance Computing (HPC) center of the University of at each time step have been set so as to guarantee a cumulative
Modena and Reggio Emilia. Two different grids are employed for mass conservation error below 10−12 .
For the proposed methodology, a value of 7.2 μs/DoF is mea-
sured. This is perfectly comparable to the most up-to-date codes
being developed in recent years, see for example Table 4 in Ref.
[29] where a cost of 17.6 μs is reported as the cpu time per DoF
measured for a 8.9 billion point Cartesian mesh. Table 1 shows that
despite the overhead represented by the unused cells lying outside
the cylindrical boundary (about 25% of the total number of vol-
umes, see Fig. 13(a)), the present method is more than five times
faster than the conventional unstructured solver considered.
Table 1
Computational performance of the proposed method, as compared
with a conventional unstructured Finite Volume solver, on the tur-
bulent pipe flow case with Lx = 10R at Reτ = 180. The number of
cells reported in parentheses for the present method denotes the
actual number of elements lying in the fluid region.
N. of cells ×10−6 y+
w,mean x+ μs/DoF
present 16.8 (12.7) 0.3 7 7.2
Fig. 12. Speedup S and number of computational cores N for the current imple-
OpenFOAM® 13.6 0.5 7 40.2
mentation of the proposed technique.
158 D. Angeli and E. Stalio / Computers and Fluids 183 (2019) 148–159
Fig. 13. Details of the grids employed for testing performances on the turbulent pipe flow case at Reτ = 180: (a) Cartesian grid with immersed boundaries (present method);
(b) Boundary-fitted grid (OpenFOAM).
where ci,j,k are the Fourier coefficients of c(x, y, z) and the Poisson [6] Leveque R. High resolution finite volume methods on arbitrary grids via wave
Eq. (47) becomes propagation. J Comput Phys 1988;78(1):36–63.
[7] Mittal R, Iaccarino G. Immersed boundary methods. Annu Rev Fluid Mech
2005;37:239–61.
− φi jk π 2 i2 + j2 + k2 eiπ ix eiπ jy eiπ kz = ci jk eiπ ix eiπ jy eiπ kz [8] Bonfigli G. High-order finite-difference implementation of the im-
i, j,k i, j,k mersed-boundary technique for incompressible flows. Comput Fluids
2011;46(1):2–11.
(49) [9] Gautier R, Biau D, Lamballais E. A reference solution of the flow over a circular
cylinder at Re=40. Comput Fluids 2013;75:103–11.
given the uniqueness of the Fourier expansion [10] Bernardini M, Modesti D, Pirozzoli S. On the suitability of the immersed
ci jk boundary method for the simulation of high-reynolds-number separated tur-
φi jk = − (50) bulent flows. Comput Fluids 2016;130:84–93.
π 2 i2 + j 2 + k2 [11] Gronskis A, Artana G. A simple and efficient direct forcing immersed bound-
ary method combined with a high order compact scheme for simulating flows
Eq. (50) is used in the computational code implemented in order to with moving rigid boundaries. Comput Fluids 2016;124:86–104.
[12] Frisani A, Hassan Y. On the immersed boundary method: finite element versus
decouple the three dimensional Poisson equation in one direction
finite volume approach. Comput Fluids 2015;121:51–67.
(or more directions) for domain decomposition purposes. [13] Krause D, Kummer F. An incompressible immersed boundary solver for mov-
ing body flows using a cut cell discontinuous Galerkin method. Comput Fluids
∂ 2 φi ∂ 2φ
(y, z ) + 2 i (y, z ) = ci (y, z ) + φi (y, z )π 2 i2 (51) 2017;153:118–29.
∂y 2 ∂z [14] Barozzi GS, Bussi C, Corticelli MA. A fast cartesian scheme for unsteady heat
diffusion on irregular domains. Numer Heat Trans B 2004;46:56–77.
The algorithm for decoupling a three-dimensional Poisson problem [15] Piller M, Polidoro S, Stalio E. Multiplicity of solutions for laminar, fully-devel-
in nx two-dimensional problems can be sketched as follows oped natural convection in inclined, parallel-plate channels. Int J Heat Mass
Trans 2014;79:1014–26.
• Compute the forward Fourier transform along direction x of the [16] Stalio E, Piller M. Direct numerical simulation of heat transfer in converging–
diverging wavy channels. J Heat Trans 2007;129(7):769–77.
right hand side ci,j,k for all indices j, k [17] Piller M, Stalio E. Numerical investigation of natural convection in inclined
• Solve plane by plane a two dimensional Poisson problem where parallel-plate channels partly filled with metal foams. Int J Heat Mass Trans
the right hand is the DFT along x of ci,j,k and the solution matrix 2012;55(23–24):6506–13.
[18] Versteeg HK, Malalasekera HK. An introduction to computational fluid dynam-
is obtained from subtraction of the quantity (iπ /lx )2 from the ics. Harlow: Longman Scientific and Technical; 1995.
diagonal of the two-dimensional solution matrix [19] Piller M, Stalio E. Compact finite volume schemes on boundary-fitted grids. J
• The solution to the three dimensional Poisson problem is ob- Comput Phys 2008;227(9):4736–62.
[20] Patankar SV. Numerical heat transfer and fluid flow. 1st ed. McGraw-Hill; 1980.
tained by the backward Fourier transform along direction x of [21] Chorin AJ. Numerical solution of the Navier-Stokes equations. Math Comput
the plane-by plane solution. 1968;22:754–62.
[22] Gresho PM, Sani R. On pressure boundary conditions for the Navier-Stokes
equations. Int J Numer Methods Fluids 1987;7:1111–44.
A.2.2. Solution to the Poisson problem in each plane [23] Rehme K. Convective heat transfer over rod bundles. Handbook of Single-Phase
As described in Section A.2.1, the domain decomposition along Convective Heat Transfer. Kakac S, Shah RK, Aung W, editors. John Wiley and
the homogeneous direction x decouples the three dimensional Sons; 1987.
[24] Eggels J, Westerweel J, Nieuwstadt F. Fully developed turbulent pipe flow: a
Poisson problem in nx two dimensional problems. Each of these comparison between direct numerical simulation and experiment. J Fluid Mech
is solved by a Matrix Decomposition technique which is derived 1994;268:175–209.
from Ref. [33]. The Matrix Decomposition technique employed per- [25] Piller M. Direct numerical simulation of turbulent forced convection in a pipe.
Int J Numer Meth Fluids 2005;49.
forms the diagonalization of suitable sub-matrices of the global
[26] Hunt JCR, Wray AA, Moin P. Eddies, stream, and convergence zones in turbu-
matrix that couple the problem in two directions. After the Poisson lent flows. Center For Turbulence Research; Report CTR-S88; 1988.
equation is further decoupled the Poisson problem is finally de- [27] Weller H, Tabor G, Jasak H, Fureby C. A tensorial approach to computa-
tional continuum mechanics using object-oriented techniques. Comput Phys
composed in a set of nx × ny one dimensional cyclic Poisson prob-
1998;12(6):620–31.
lems. These are discretized with standard Finite Volume techniques [28] Issa R. Solution of the implicitly discretised fluid flow equations by operator-s-
within the second order accuracy and the cyclic, tridiagonal lin- plitting. J Comput Phys 1986;62(1):40–65.
ear problems derived are conveniently solved by the Thomas algo- [29] Borrell G, Sillero J, Jiménez J. A code for direct numerical simulation of turbu-
lent boundary layers at high reynolds numbers in BG/p supercomputers. Com-
rithm. put Fluids 2013;80(1):37–43.
[30] Gresho PM. On the theory of semi-implicit projection methods for incompress-
References ible flow and its implementation via a finite element method that also intro-
duces a nearly consistent mass matrix. part 1: theory. Int J Numer Methods
[1] Lee M, Moser RD. Direct numerical simulation of turbulent channel flow up to Fluids 1990;11:587–620.
Reτ ≈ 5200. J Fluid Mech 2015;774:395–415. [31] Laizet S, Li N. Incompact3d: a powerful tool to tackle turbulence prob-
[2] Yokokawa M, Itakura K, Uno A, Ishihara T, Kaneda Y. 16.4Tflopss direct nu- lems with up to o(105 ) computational cores. Int J Numer Methods Fluids
merical simulation of turbulence by a fourier spectral method on the earth 2011;67(11):1735–57.
simulator. In: Proceedings of the ACM/IEEE SC Conference; 2002. p. 1–17. [32] Lamballais E, Silvestrini J, Laizet S. Direct numerical simulation of a separa-
[3] Peskin C. The immersed boundary method. Acta Numer 2002;11:479–517. tion bubble on a rounded finite-width leading edge. Int J Heat Fluid Flow
[4] Leveque RJ, Li Z. Immersed interface method for elliptic equations with 2008;29(3):612–25.
discontinuous coefficients and singular sources. SIAM J Numer Anal [33] Babu V, Korpela S. On the direct solution of Poisson’s equation on a non-uni-
1994;31(4):1019–44. form grid. J Comput Phys 1993;104(1):93–8.
[5] Patankar SV. A numerical method for conduction in composite materials, flow
in irregular geometries and conjugate heat transfer. In: Proceedings of the 6th
International Heat Transfer Conference, 3; 1978. p. 297.