# The Kelvin-Helmholtz instability in two dimensions

On this page a classical shear instability is simulated. According to the Kelvin-Helmholtz instability description, a thin shearing layer may be unstable and can then roll-up on itself creating vortices, or so-called Kelvin-Helmholtz billows. On this page the effect of the fluid’s viscosity on the observed dynamics is investigated.

## Set-up

A fluid with an infinetly thin shear layer is initialized. The difference in velocity between the induvidual layers U and the length scale associated with the periodicity of the solution (\mathcal{L}) in our set-up can be used to normalize the value of the fluid’s viscosity. Resulting in a Reynolds number,

\displaystyle Re=\frac{U\mathcal{L}}{\nu}.

In this study, the Reynolds number is varied to be Re=\{20000, 40000, 80000\}. In order to analyze the flow’s evolution, the emergence of coherent vortex structures is monitored. Therefore, the number of vortices is counted using the marvalous tag() function. It is remarkable how complex seamingly simple things can be.

#include "navier-stokes/centered.h"
#include "tag.h"

int maxlevel;
double vis, Re;
FILE * fpn;
char fnamev[99];
char fnameg[99];

As mentioned earlier periodic boundaries are used in the stream-wise direction. Furthermore, the boundaries in the span-wise direction of the square domain are of the free-slip type.

int main(){
periodic(left);
X0 = Y0 = -L0/2;

A loop is used to run the simulation for the three different Reynolds numbers, increasing the grid resolution each iteration to a maximum of 11 levels for Re = 80000.

  maxlevel = 9;
for (Re = 20000; Re <= 80001; Re = Re*2.){
init_grid (1 << 4);
run();
maxlevel++;
}
}

event init(t = 0){
vis = 1./Re;
const face vector muc[] = {vis, vis};
mu = muc;

The shear layer is initialized and a small random perturbation is added to the span-wise-velocity-componenent field to kick-off the growth of the instability.

  astats adapting;
do{
foreach()
u.x[] = 0.5 - (y < 0);
boundary({u.x});
foreach()
u.y[] = 0.004*noise();

Output file names are defined and their names carry a Reynolds-number identifier.

  sprintf (fnamev, "KH%g.mp4", Re);
sprintf (fnameg, "KHlev%g.mp4", Re);
char name1[99];
sprintf (name1, "nrofvortices%g", Re);
fpn = fopen (name1, "w");
}

## Output

The usual suspects of animation types are generated. Meaning that the evolution of the vorticity (\omega) field and the grid resolution are visualized. The maximum value of the vorticity field (\|\omega\|_{max}) is monitored. consequently, a vortex can be defined as a connected region with a vorticity value \|\omega\|>\|\omega\|_{max}/3. Based on this definition, we count and log the number of vortices.

event output (t = 0.005; t += 0.01){
scalar omega[], lev[], f[];
int n = 0;
double m = 0;
foreach(reduction(+:n) reduction(max:m)){
n++;
lev[] = level;
omega[] = (u.x[0,1]-u.x[0,-1] - (u.y[1,0]-u.y[-1,0]))/(2*Delta);
if (fabs(omega[]) > m)
m = fabs(omega[]);
}
boundary({omega});
output_ppm (omega, file = fnamev, n = 512, min = -10, max = 10, linear = true);
output_ppm (lev, file = fnameg, n = 512, min = 2, max = 11);
foreach()
f[] = (omega[] > m/3.); //1 or 0
int nrv = tag(f);
fprintf (fpn,"%g\t%d\t%g\t%d\t%d\t%d\n",t, nrv, m,
n,((1 << (maxlevel*dimension)))/(n), i);
}

## The adaptation and the end-of-run events

The grid resolution is adapted based on the wavelet-estimated discretization error in the representation of the velocity vector field. In the end-of-run event, the files associated with the run’s output file pointer is closed.

event adapt (i++)

event end (t = 5.)
fclose(fpn);

## Results

One may visually inspect the evolution of the various flow set-ups using the visualizations of the solutions and their grids.

For Re=20000,

Evolution of the vorticity field

Evolution of the grid structure

Re=40000,

Evolution of the vorticity field

Evolution of the grid structure

and Re=80000,

Evolution of the vorticity field

Evolution of the grid structure

It seems as if the fastest growing mode is a function of the Reynolds number, this is consistent with the linear perturbation analysis that is presented in many text books on this topic. The ratio of the wavelength associated with this mode and the periodicity length scale (\mathcal{L}) is related to the number of vortices that emerge. We can plot the evolution of the number of vortices for the the different runs.

set yr [1:50]
set xlabel 't U/L [-]'
set ylabel 'Vortex Count [-]'
set logscale y
set size square
plot 'nrofvortices20000' u 1:2 w l lw 3 t 'Re=20000' ,\
'nrofvortices40000' u 1:2 w l lw 3 t 'Re=40000' ,\
'nrofvortices80000' u 1:2 w l lw 3 t 'Re=80000'

Apart from the fact that the used vortex detection algorithm intermittently identifies decaying shear layers as vortices, these numbers seem to correspond with what we saw in the visualizations. Well done tag() function!

Finally, we check if it was sensible to use an adaptive grid by plotting the grid-cell-compression ratio (\Pi) over time. This ratio is defined as the number of grid cells required to fill the domain with an equidistant and static grid with a cell size \Delta_{min} divided by the number of grid cells employed by the adaptation algoirithm (i.e. \Pi \geq 1).

set yr [1:5000]
set xlabel 'Solver Iteration'
set ylabel 'Pi'
set key left
plot 'nrofvortices20000' u 6:5 w l lw 3 t 'Re=20000' ,\
'nrofvortices40000' u 6:5 w l lw 3 t 'Re=40000' ,\
'nrofvortices80000' u 6:5 w l lw 3 t 'Re=80000'