MESH – FREE COMPUTATIONAL FLUID DYNAMICS:

A RADIAL BASIS FUNCTION COLLOCATION METHOD APPROACH

Božidar Šarler

Laboratory for Multiphase Processes, Nova Gorica Polytechnic, Vipavska 13, SI-5000 Nova Gorica, Slovenia
bozidar.sarler@p-ng.si

ABSTRACT
This paper explores the application of the mesh-free radial basis function collocation method for solution of heat
transfer and fluid flow problems. The solution procedure is represented for a Poisson reformulated general transport
equation in terms of a-symmetric, symmetric and modified (double consideration of the boundary nodes) collocation
approaches. In continuation, specifics of a primitive variable solution procedure for the coupled mass, momentum,
and energy transport representing the natural convection in an incompressible Newtonian Bussinesq fluid are
elaborated. A comparison of different collocation strategies is performed based on the two dimensional De Vahl
5 6
Davis steady natural convection benchmark with Prandtl number Pr =0.71, and Rayleigh numbers Ra = 10 , 10 .
Multiquadrics radial basis functions are used. The three methods are assessed in terms of streamfunction extreme,
cavity Nusselt number, and midplane velocity components. Best performance is achieved with the modified
approach.

INTRODUCTION the numerical solution of Eulerian transport equation,

The development of efficient as well as simple
algorithms for the numerical solution of partial
differential equations (PDEs) is of major interest in
applied sciences and engineering. The most popular
discrete approximate methods for PDEs are nowadays
the finite difference (FDM), finite volume (FVM), the
finite element (FEM), the spectral (SM), and the
boundary element methods (BEM). Despite the powerful
features of these methods, there are often substantial
difficulties in applying them to realistic, geometrically
complex three dimensional transient situations. A
common drawback of the mentioned methods is the need
to create a polygonisation, either in the domain and/or on
its boundary. This type of meshing is often the most time
consuming part of the solution process and is far from
beeing fully automated.
In recent years, a new class of methods has been
developed which do not require polygonisation but use a
set of nodes to approximate the solution. The rapid
development of these type of methods and their
classification is elaborated in the very recent
monographs [1,2]. A most simple class of mesh-free
methods in development today are the Radial Basis
Function [3] Collocation Methods (RBFCM).
GOVERNING EQUATIONS
For the present purposes, a transport phenomena
problem can be briefly described in a general manner as
defined on a fixed domain Ω with boundary Γ , of the
kind
∂ r
 ρ f ( Φ )  + ∇ ⋅  ρ vf ( Φ )  = ∇ ⋅ ( D∇Φ ) + S (1)
∂t
r
with ρ , Φ , t , v , S , and D standing for the density,
transport variable, time, velocity, source, and diffusion
matrix. The transport variable stands, for instance for
the velocity component in each co-ordinate direction, or
temperature, or the mass fraction of a chemical species.
The function f denotes the relation between the
transported and the diffused variable such as for example
the relation between the enthalpy and the temperature.
The solution of the governing equation for the
dependent variable Φ at final time t = t0 + ∆t is sought,
where t0 represents the initial time and ∆t the positive
time increment. The solution is constructed by the initial
and boundary conditions that follow. The initial value of
the transport variable Φ at a point with position vector
r
p and time t0 is defined through the known function Φ 0
r
Φ ( p , t0 ) = Φ 0 ; p ∈ Ω ∪ Γ (2)
The boundary Γ is divided into not necessarily
connected parts Γ and Γ with Dirichlet and Neumann
D N RADIAL BASIS FUNCTION COLLOCATION
boundary conditions, respectively. These boundary METHOD
r r
conditions are at the boundary point p with normal nΓ
Description of fields and partial derivatives
and time t < t ≤ t0 + ∆t defined through known The unknown filed Φ is calculated in
r
functions Φ and Φ
D N points pn ; n = 1,..., N = N Γ + N Ω .The first NΓ belong
to the boundary and the last NΩ to the domain. The
r ∂ r unknown filed is approximated by M approximation
Φ = ΦD ; p ∈ ΓD , r Φ = Φ ; p∈Γ
N N
(3)
∂n Γ functions ψ m and their coefficients ς m of the type

The involved parameters of the governing equation M

r r
and boundary conditions are assumed to depend on the Φ ( p ) ≈ ∑ ς mψ m ( p ) (12)
transport variable, space and time. The solution n =1
procedure thus inherently involves iterations. The
governing equation is transformed into a Poisson form as Similarly, a partial derivative over coordinate ζ m of point
follows. The diffusion matrix is split into constant r
isotropic part DI and the remaining an-isotropic p (ζ 1 , ζ 2 , ζ 3 ) can be approximated as
part D '
∂ r M
∂ r
Φ ( p) ≈ ∑ς m ψ m ( p )  (13)
D = DI + D ' (4) ∂ζ i n =1 ∂ζ i

The transport equation is subsequently cast into Poisson

∂2 r M
∂2 r
form Φ ( ) ∑ m 2 ψ m ( p ) 
p ≈ ς (14)
∂ζ i2
n =1 ∂ζ i
r
∇2Φ = θ + ∇ ⋅ Θ (5)
by assuming sufficient admissibility of functions ψ m .
with A suitable function is the multiquadric

∂ 
θ =   ρ f ( Φ )  − S  / D
r
( )
1/ 2
(6) ω m ( p ) = rm2 + c 2 (15)
 ∂t 
r r with rm standing for the Euclidean distance between the
Θ =  ρ vf ( Φ ) − D ' ∇Φ  / D (7) r r
field point p and the reference point pm and c for the
The partial time derivative can be approximated by a free parameter. The coefficients ς m of the solution can
two-level fully implicit finite difference be determined in various ways. The following three
collocation strategies have been put into context of
∂ 1 solving transport phenomena in this paper.
 ρ f ( Φ )  ≈  ρ f ( Φ ) − ρ0 f ( Φ 0 )  (8)
∂t ∆t
A-symmetric collocation
The inhomogenous terms are Taylor expanded as The most simple is the a-symmetric collocation,
originally proposed by Kansa [4,5]. The coefficients ς m
r r r
θ = θ + θ ,Φ ( Φ − Φ ) , Θ = Θ + Θ,Φ ( Φ − Φ ) (9) of the solution are determined from a system of
collocation equations. The first NΓ equations read
with ‘bar’ denoting value at previous iteration. The final
form of the transformed equation, suitable for iterative
NΓ
∂
solution than becomes ∑ς
n =1
m ( χ iDψ mi + χ iD
∂nΓ
ψ mi ) = χ iD Φ iD + χ iN Φ iN

(16)
∇ 2Φ = Q ( Φ ) (10)
r r and the last M − NΓ equations read
Q ( Φ ) = θ + θ ,Φ ( Φ − Φ ) + ∇ ⋅ Θ + ∇ ⋅ Θ,Φ ( Φ − Φ )
M
(11)
∑
n = NΓ +1
ς m∇ 2ψ mi = Qi (17)
by setting r
is satisfied in all gridpoints pi .
ψ mi = ω mi , M =N (18) The steady state criterion
The steady-state is assumed to be reached when the
The first NΓ equations have been written for the conditon
Ω
boundary points, and the last N for the domain points r r
r | Φ j +1 ( pi , t + ∆t ) − Φ ( pi , t ) | < ε ste (23)
≡ F ( pi ) , where F stands for an arbitrary
and Fi = r
function. The introduced boundary condition indicators is satisfied in all gridpoints pi .
take form
NATURAL CONVECTION PROBLEM
r r
1; pi ∈ Γ D 1; pi ∈ Γ N
χi =  r
D
D ,
χi =  r
N
(19) Natural convection
0; pi ∉ Γ 0; pi ∉ Γ
N
In order to demonstrate the application and suitability
of the represented solution procedures for a typical
Symmetic collocation computational fluid dynamics problem, the following
The system of equations (16,17) can be made system of coupled mass, momentum, and energy
symmetric [6] by setting the Fasshauer’s ansatz equations is used. It describes the natural convection of
an incompressible Newtonian Bussinesq fluid

NΓ
∂ω mi  M r
ψ mi = ∑  χ mDω mi + χ mN  ∑ ∇ ω mi ,
+ 2
∇⋅v = 0 , (24)
m =1  ∂nΓ  m = NΓ +1
M =N ∂ r rr
(20)
( ρ v ) + ∇ ⋅ ( ρ vv ) =
∂t
( )
The symmetric solution gives better results [7] as the a- r r
symmetric one. −∇P + µ∇ 2v + ρ a 1 − β T − T0  , (25)

Modified collocation
If we attempt the solution only in the boundary nodes
or only in the domain nodes the information either on the
governing equation either or on the boundary conditions
is lost in these two extremes. In boundary points, the
governing equation as well as boundary conditions have
to be valid. Respectively, the system of equations
(16,17) can be made symmetric and consistent by
attempting the solution with double consideration of the
boundary nodes, as deduced by Chen and Tanaka [8], by
setting
∂ r
( ρ cPT ) + ∇ ⋅ ( ρ vcPT ) = k∇ 2T (26)
∂t
r
with P standing for pressure, µ for viscosity, a for
acceleration, β for the volumetric thermal expansion
coefficient, T0 for the Bussinesq reference temperature,
cP for specific heat, k for thermal conductivity, and
T for temperature. All material properties are constant.

Solution strategy
NΓ
 ∂ω mi  M
ψ mi = ∑  χ mDω mi + χ mN  ∑ ∇ ω ( m − N Γ )i ,
+ 2 The energy equation is coupled with the momentum
∂nΓ  m = NΓ +1 equation through the velocity field and the momentum
m =1 
equation is coupled with the energy equation through the
M = NΓ + N (21) body force. Respectively, the solution inherently
involves iterations. Let us assume the velocity, pressure
Collocation is first made in the boundary nodes where and temperature fields are all known at iteration level
the boundary conditions are valid and subsequently in j .What follows explains the solution at the iteration
the boundary and domain nodes where the governing level j + 1 .
equation is valid. By inserting the approximation (21)
into framework (16,17) gives an ( N Γ + N ) × ( N Γ + N ) Solution of the momentum equation: The solution of the
symmetric system of algebraic equations. momentum equation at iteration level j + 1 is obtained
in the following way: The Pressure Poisson Equation
The convergence criterion
(PPE) is constructed by taking the divergence of the
The timestep iterations are stopped when the condition
momentum equation
r r
| Φ j +1 ( pi , t ) − Φ j ( pi , t ) | < ε itr (22)
  1 conditions are valid
∇ 2 P j +1 = ∇ ⋅ −
 ∆t
r
(r r r
)
ρ v j − ρ v0 − ∇ ⋅ ρ v j v j ( )
∂ % j +1 r
P = 0; p ∈ Γ (33)
∂nΓ
r r
(
+ µ∇ v j + ρ a 1 − β T j − T0  (27)
2
)}
Both PPE and PCPE are singular due to the presence
The Neumann pressure boundary conditions can be of the Neumann boundary conditions over the whole
r
defined along the whole boundary Γ by taking the boundary. One of the domain grid-points p0 is fixed to
scalar product of the momentum equation with the
the reference pressure P0 in case of the PPE, and to 0 in
normal on the boundary. This gives
case of the PCPE, in order to avoid the singularity
∂ j +1  1
∂nΓ
P = −
 ∆t
r
( r
)
r r
ρ v j − ρ v0 − ∇ ⋅ ρ v j v j ( ) r r
P j +1 ( p0 ) = P0 ; p0 ∈ Ω (34)
r r
P% j +1 ( p ) = 0; p ∈ Ω
( )}
r r r (35)
+ µ∇ 2 v j + ρ a 1 − β T j − T0  ⋅ nΓ (28)
0 0

After calculating the pressure gradient field the velocity
field at iteration level j + 1 can be solved from the
velocity Poisson equation
r
The domain point which coincides with the point p0
is treated in a formally equivalent way as it would
represent a boundary point with prescribed Dirichlet
boundary conditions.

r r
11
∇ 2vˆ j +1 = 
µ  ∆t
( r r r
)
ρ vˆ j +1 − ρ v0 + ∇ ⋅ ρ v j v j ( )
Solution of the energy equation: The energy equation is
solved from

1 1
r
(
+∇P j +1 − ρ a 1 − β T j − T0  (29) )} ∇ 2T j +1 = 
k  ∆t
(
ρ cPT j +1 − ρ cPT0 )
The “hat” on the velocity denotes that the velocity does
r
}
+∇ ⋅ ( ρ v j +1cPT j ) (36)
not correspond to the mass conservation in general. The
incompressibility is enforced through the pressure P% and At the end of the iteration, the body force is updated
r with the new value of the temperature and the solution is
velocity corrections v% , which ensure repeated until the iteration and steady state conditions
r r are met for the pressure, velocity components, and
r
(
∇ ⋅ v j +1 = ∇ ⋅ vˆ j +1 + v% j +1 ) (30) temperature.

NUMERICAL RESULTS
Consider that the velocity corrections occur
exclusively due to action of the pressure correction P% Natural convection [9] benchmark
Geometry is a square with dimension L . Two
ρ r j +1 dimensional Cartesian coordinates px , p y are used. The
ε Pv v% = −∇P% j +1 (31)
∆t square is extending from px − ≤ px ≤ px + ,
p y − ≤ p y ≤ p y + , pxy ± = ± L / 2 . Upper and lower
where ε Pv represents a heuristic velocity correction –
boundaries are insulated, the left boundary is subject to
pressure correction relaxation factor. The pressure
correction can thus be calculated from the velocity field temperature T+ , the right boundary is subject to
r
v̂ through the Pressure Correction Poisson Equation temperature T− . The solution is performed for Prandtl
(PCPE) number Pr = µ cP / k = 0.71 , and for Rayleigh
numbers Ra = a β (T+ − T− ) L Pr ρ / µ = 10 , 10 .
3 2 2 5 6
ρ r
∇ 2 P% j +1 = ε Pv vˆ (32) The square is discretized in an uniform grid
∆t
20 × 20 or 30 × 30 . The first(second) grid involves 441
deduced from the equations (30) and (31). Since no (961) points of which 80(120) are on the boundary and
correction is needed in the direction normal to the solid 361(841) in the domain. The dimensionless time-step
boundary, the following pressure correction boundary used in the calculations has been set to ∆t =10 / Ra .
The constant in multiquadrics has been set to p y+
1 M
∂
c =1.75∆L , where ∆L represents a typical grid Nu =
T+ − T−
∑ς ∫
m =1
T
m
∂p
ψ m dp y (42)
distance. The velocity correction – pressure correction py− x

relaxation factor has been set to ε Pv =1 . The iteration

criterion and the steady-state criterion have been set to All involved integrals of RBFs in numerical evaluation
of the Ψ and Nu are evaluated in an analytical way.
ε itr =10−5 , ε ste =10−4 , respectively. The temperature is
distributed linearly from the left to the right of the cavity Numerical implementation
and no flow is assumed at the initial time. The numerical implementation is made double
The simulations have been performed with three precision in Compaq Visual Fortran 90 with IMSL
different solution procedures: with the a-symmetric one library. Test cases have been run on an HP Omnibook
(ASY) [10], with the symmetric one (SYM) and with the XE3 laptop with an Intel Pentium III 850MHz
modified one (MOD). The dimensions of the Poisson
processor, 256MB ram and Windows 2000 operating
equations are for the grids 20 × 20 equal to 441× 441 in
system. The system matrices have been first LU
case of the ASY and SYM procedures, and equal to
decomposed by using the routine DLFRGT and then
521× 521 in case of the MOD procedure. The solved by the routine DLFSRG at each iteration. The
dimensions of the Poisson equations are for the
30 × 30 MOD CPU time is 7.5 hours for Ra = 106 .
grids 30 × 30 equal to 961× 961 in case of the ASY and
SYM procedures, and equal to 1081× 1081 in case of Table 1: The number of time-steps, and internal time-
the MOD procedure. step iterations used as a function of the solution
procedure and problem.
Calculation of the streamfunction procedure time-steps iterations
Ra
The streamfunction Ψ is calculated as
105 ASY 188 2062
py px 105 SYM 191 1991
Ψ= ∫ v dp
py−
x y , Ψ=− ∫ v dp
px −
y x (37) 105 MOD 194 2007
106 ASY 295 3531
106 SYM 299 3501
A variation of the velocity components over the domain
Ω and boundary Γ is based on the global approximation 106 MOD 304 3529

M M
vx = ∑ ς mvxψ m , v y = ∑ ς myψ m
v
(38) Table 2: Natural convection in a square cavity. The
m =1 m =1
maximum value of the horizontal velocity vx max on the
Respectively, Ψ can be calculated as vertical midplane.
Ra grid ASY SYM MOD Ref.[9]
M py M px
10 5 20 × 20 66.60 66.90 67.03 68.59
Ψ = ∑ ς mvx ∫ ψ m dpy , Ψ = −∑ ς m ∫ψ
v
y
dpx (39) 30 × 30 67.59 68.01 68.27 68.59
m =1 py− m =1 px −
m
105
106 20 × 20 205.11 207.67 209.31 219.36
Calculation of the Nusselt number 106 30 × 30 211.67 213.81 215.56 219.36
The cavity Nusselt number Nu is calculated as

py+ Table 3: Natural convection in a square cavity. The

1 ∂
Nu =
T+ − T− ∫p ∂px Tm dpy (40) maximum value of the vertical velocity v y max on the
y− horizontal midplane.
Ra grid ASY SYM MOD Ref.[9]
A variation of the temperature over the domain Ω 5 20 × 20 31.86 31.96 32.11 34.73
and boundary Γ is based on the global approximation
10
105 30 × 30 32.51 33.70 34.08 34.73
M
106 20 × 20 59.67 61.01 61.48 64.63
T = ∑ ς mTψ m (41) 30 × 30 61.55 62.64 63.12 64.63
m =1 106
Respectively, Nu can be calculated as
Table 4: Natural convection in a square cavity. Values of REFERENCES
stream-function extreme Ψ at the center of the cavity as
a function grid density and collocation scheme. 1. S.N. Atluri and S. Shen, The Meshless Method,
Ra grid ASY SYM MOD Ref.[9] Tech Science Press, Encino, 2002.
5 20 × 20 8.861 8.880 8.901 9.111 2. G.R. Liu, Mesh Free Methods, CRC Press, Boca
10 Raton, 2003.
105 30 × 30 8.907 8.950 9.073 9.111 3. M.D. Buhmann, Radial Basis Functions, Acta
106 20 × 20 15.03 15.41 15.88 16.32 Numerica, Cambridge University Press,
Cambridge, pp.1-38, 2000.
106 30 × 30 15.91 16.07 16.22 16.32
4. E.J. Kansa, Multiquadrics – a Scattered Data
Approximation Scheme with Application to
Table 5: Natural convection in a square cavity. Values of Computational Fluid Dynamics, Part I,
cavity Nusselt number Nu as a function of grid density Comp.Math.Applic., vol.19, pp.127-145, 1990.
and collocation scheme. 5. E.J. Kansa, Multiquadrics – a Scattered Data
Ra grid ASY SYM MOD Ref.[9] Approximation Scheme with Application to
20 × 20 4.647 4.640 4.573 4.519 Computational Fluid Dynamics, Part II,
105 Comp.Math.Applic., vol.19, pp.147-161, 1990.
105 30 × 30 4.530 4.528 4.523 4.519 6. G.E. Fasshauer, Solving Partial Differential
106 20 × 20 8.921 8.909 8.876 8.800 Equations by Collocation with RBFs,
Proc.Symp.Chamonix, pp.1-8, 1996.
106 30 × 30 8.869 8.845 8.834 8.800
7. H. Power and W. Barraco, A Comparison Analysis
Between Unsymmetric and Symmetric RBFCMs
Discussion of the results for the Numerical Solution of PDF’s,
The number of required iterations remain almost the Comp.Math.Applic., vol.43, pp.551-583, 2002.
same in all three methods as seen in Table 1. Without 8. W. Chen and M. Tanaka, New Insights into
exception, the SYM approach gives better results as the Boundary-Only and Domain-Type RBF Methods,
ASY one, and the MOD approach gives better results as J.Nonlin.Sci., vol.1, pp.145-151, 2000.
the SYM one, as could be clearly concluded from the 9. G. De Vahl Davis, Natural Convection in a Square
Tables 2-5. Cavity, A Bench Mark Solution, Int.J.Numer.
Methods Fluids, vol.3, pp.249-264, 1983.
CONCLUSIONS 10. B. Šarler, J.Perko, C.S.Chen and G.Kuhn, A
Meshless Approach to Natural Convection, CD
This paper shows three RBFCM strategies for dealing ROM Proc.Symp.ICCES-01, 6 pages, 2001.
with transport phenomena. The asymmetric, symmetric 11. B. Šarler, J.Perko and C.S. Chen, Radial Basis
and modified formulations are critically evaluated based Function Collocation Method Solution of Natural
on the classical De Vahl Davis benchmark test. Even the Convection in Porous Media, Int.J.Numer.Methods
most simple of them, the a-symmetric formulation Heat & Fluid Flow, (in print).
performs well for a broad spectra of CFD problems [10, 12. I.Kovačević, A.Poredoš and B.Šarler, Solving the
11,12]. The other two formulations can be successfully Stefan problem by the RBFCM, Numer.Heat
applied to boost the performance of the first one on the Transfer, (in print).
expense of slightly more complicated coding. The main 13. C.S. Chen, C.A. Brebbia and H. Power, Dual
disadvantage of the presented method represent full Reciprocity Method Using Compactly
matrices that are difficult and expensive to solve for Supported RBFs, Comm.Numer.Methods Eng.,
large problems. This issue might be in the future
vol.15, pp.137-150, 1999.
mitigated or overcome by the use of the compactly
14. G.E. Fasshauer, Solving Differential Equations with
supported radial basis functions [13], multilevel radial
RBFs: multilevel methods and smoothing,
basis functions [14], iterative solvers [15], adaptive grid
Adv.Comput.Math., vol.11, pp.139-159, 1998.
or domain decomposition [16]. It would be too
15. V. Bulgakov, B.Šarler and G.Kuhn, Iterative
ambitious to claim that the represented method can cope
solution of systems of equations in the dual
a wide variety of large-scale computational science and
reciprocity boundary element method for the
engineering problems at this point. Additional research
diffusion equation, Int.J.Numer.Methods Eng.,
in this direction is definitely required.
vol.43, pp.713-732, 1998.
ACKNOWLEDGEMENT
16. N. Mai-Dui and T. Tran-Cong, Mesh free radial
basis function network methods with domain
decomposition for approximation of functions and
A part of the present research has been supported by a
numerical solution of the Poisson’s equation,
NATO grant No.PST.CLG.9977633 and a grant of the
Eng.Anal.Bound.Elem., vol.26, pp.133-156, 2002.
Slovenian MESS No.L2-5387-1540-03.