Professional Documents
Culture Documents
Department of Mechanical and Materials Engineering, Queen’s University, Kingston, ON, K7L 3N6, Canada
Email: pollard@me.queensu.ca
The Boltzmann equation is difficult to solve for two Once the fi ’s are known, the hydrodynamic quantities
reasons: the single particle distribution is a function of can be computed by the following summations:
W7−18 = 1/36
N−1
ρ(x,t) = ∑ fi (x,t) (7) (10)
i=0
3 S IMULATION D ETAILS
One major drawback of the LBM is that currently
it can only be used with a uniform grid. This
means that the grid resolution must be the same in
all three coordinate directions. In a simulation of
Figure 1: D3Q19 lattice velocities.
turbulent channel flow the grid resolution required in
The equilibrium distribution for the LBM is based on the wall-normal direction is much higher than that
the Maxwell distribution, which is the equilibrium dis- needed in the stream-wise and span-wise directions.
tribution for the Boltzmann equation. To generate the Ideally, a stretched grid would be implemented to
LBM equilibrium distribution function, a low-Mach exploit this feature to make the simulation more
number assumption is made. This assumption allows computationally efficient. Since this is currently not
the Maxwell distribution to be expanded as a Taylor possible with the LBM, the computational cost of
series about zero velocity. The terms up to second or- simulating a channel flow using the LBM is much
der in u are retained to give: greater than that of a simulation using a spectral
method. However, turbulent channel flow is an
ideal simulation to run in order to benchmark the
accuracy of the LBM for wall bounded turbulent flows.
3(ci · u) 9(ci · u)2 3u2
(eq)
fi = Wi ρ 1 + + − (9)
c2 2c4 2c2
In this work, the minimal flow unit was simulated
where the constants Wi are the quadrature coefficients instead of the channel of [7] to reduce the computa-
(eq) tional time. The minimal flow unit restricts the size
that ensure that the moments of fi corresponding to
of the channel in both the stream-wise and span-wise
the conserved quantities are identical to the moments
directions. Due to its smaller size, the minimal
of fi . The values of these coefficients for the D3Q19
channel is not large enough to accommodate all of the
lattice are:
largest scales in the channel, and thus there are errors
induced by enforcing periodic boundaries. However,
[6] showed that, provided the channel is larger than
W0 = 1/3 a critical size, the minimal channel can sustain a
W1−6 = 1/18 turbulent flow and the near-wall turbulence statistics
agree well with the data from the full channel. By B7−18 = 1/12
simulating the minimal channel, the LBM can be (12)
implemented with a uniform grid and the simulation
can be performed in a reasonable amount of time.
The LBM code has been parallelised using MPI as
the message passing API. The code was run on a
The Reynolds number, based on the mean center- SUN Fire 15K shared-memory machine using 64
line velocity and the channel half-width was 3300. processors. The simulation completed 12 million time
The dimensions of the simulation domain were steps, corresponding to approximately 8000 large
2δ × 0.3πδ × πδ for the wall-normal, span-wise, and eddy turn-over times.
stream-wise directions respectively. The domain was
meshed with a uniform grid with 181 × 86 × 282
nodes. This corresponds to a cell size of y+ = 2,
which is approximately equal to the Kolmogorov 4 R ESULTS
length scale. The probability density values for the During the simulation, all of the relevant time averaged
model were initialised to their equilibrium values quantities needed to compute turbulence statistics up
for Poiseuille flow. In order to induce turbulence, a to second order were recorded. These included:
random noise was applied directly to the populations
for a short time using a linear congruential random
number generator. • Density (ρ)
ρ
where g is the body force vector and the weighting fac- p= (13)
tors Bi are equal to: 3
The pressure is scaled using the mean wall shear stress
and plotted relative to the pressure on the wall. The
B1−6 = 1/6 LBM simulation is within approximately 10% of the
20 run with a resolution of 1.5 wall units that showed no
appreciable change.
15
3.0
u+
10
2.5
5 2.0
u′ rms
1.5
+
0
1.0 10.0 100.0 1.0
y+ 0.5
0.0
Figure 2: LBM results for the mean velocity profile:
1.0 10.0 100.0
◦ lower wall, + upper wall, — Kim, Moser, Moin
+
(1987) [7] y
−0.4
1.5
p+
−0.6
−0.8 1.0
v′ rms
+
−1
1.0 10.0 100.0 0.5
+
y
results from [7]. Figure 5: LBM results for the wall-normal RMS veloc-
ity fluctuations: ◦ lower wall, + upper wall, — Kim,
Moser, Moin (1987) [7]
The root-mean-square velocity fluctuations in the
stream-wise, wall-normal, and span-wise directions
Currently, a simulation is running in which the length
are shown in figures 4 - 6. These figures show a
of the channel has been extended in the stream-wise
good match for the wall-normal velocity fluctuations,
direction. The new dimensions of this channel
but significant discrepancies in the stream-wise and
are 2δ × δ × 16δ, and the resolution in each of the
span-wise directions. The most significant difference
directions is 2 wall units. This channel is more than
between the two simulations is an over prediction of
four times longer than the minimal channel in the
the stream-wise velocity fluctuations by over 10%
stream-wise directions, and thus uses approximately
at the peak value. Initially, this discrepancy was
four times the number of lattice sites.
explained as a resolution problem because it was not
observed in the minimal channel simulation of [6] and
an over prediction of the peak stream-wise velocity Given the fact that this simulation contains roughly
fluctuations has been shown by [8] to be caused by four times as many computational nodes, it theoreti-
insufficient resolution. However, a simulation was cally should take four times longer to run. However,
1.5 time averaged over a much larger interval in order for
the statistics to become stationary.
u′ rms
1.5
0.8 +
1
0.6
0.5
+
u ′ v′
0.4 0
1.0 10.0 100.0
0.2 +
y
0.0
0.0 0.2 0.4 0.6 0.8 1.0 Figure 8: LBM results for the stream-wise RMS ve-
locity fluctuations (long channel simulation): ◦ lower
y+ wall, + upper wall, — Kim, Moser, Moin (1987) [7]