[tanmayagrawal7]

Hello,

I want to solve two dimensional Navier Stokes equations on a staggered grid for the case of Taylor-Green Vortex. As I am aware, we should use periodic boundary conditions on all the four boundaries. I am facing some trouble regarding implementation of this BC as I could not find enough literature about it.

Suppose, if my index ranges from 0 to N. Then, is this the correct way to use periodic BC?
u_vel(N) = u_vel(1)
u_vel(0) = u_vel(N-1)

and similarly for v_vel and pressure variables?

Thanks!

[kaya]

Yes you should use periodic bc.

I'd make it in a way that
u_vel(N+1) = u_vel(0). So my derivative matrix will be say 3-diagonal and a point at the corners.

Not that you can't do u_vel(N) = u_vel(1) but why wasting computation if you know the values

[FMDenaro]

be carefull to the link between the periodic nodes if you use staggered grid!

[tanmayagrawal7]

Originally Posted by FMDenaro

be carefull to the link between the periodic nodes if you use staggered grid!

I do not clearly get what you mean.
If I just write

u(0) = u(Nu-1), u(Nu) = u(2)
v(0) = v(Nv-1), v(Nv) = v(2)
p(0) = p(Np-1), p(Np) = p(2)

for their respective grids. Will it create some problem?

BTW, I am doing exactly this and my result diverges. I am using analytic solution as my initial conditions.

[FMDenaro]

I depends on the grid topology you are using...

For example, given L the periodicity lenght in the interval (0,L) and Nx the number of step sizes dx, the nodal grid can be written as

x(i) =i*dx i=0:Nx or
x(i)=(i-1)*dx i=1:Nx+1

now you have to define the staggering for the variables, for example in 2D

u(i,j) at x(i),y(j)-dy/2
v(i,j) at x(i)-dx/2,y(j)
p(i,j) at x(i)-dx/2, y(j)-dy/2

[sbaffini]

Usually, if you solve between 1 and N, periodicity is such that:

0 = N-1
N+1 = 2

so that:

1 = N

[tanmayagrawal7]

Originally Posted by sbaffini

Usually, if you solve between 1 and N, periodicity is such that:

0 = N-1
N+1 = 2

so that:

1 = N

Thanks for your answer!

So, as in my case of a staggered grid, I have three separate grids for U, V and P. Their grid sizes being [N,N+1],[N+1,N],[N+1,N+1] respectively.

For the U velocity, I solve NS equations for the interior points i.e. upto N-1 points in x and N points in y-direction, and similarly for the other two primitive variables. I am using an explicit finite difference based method.

So, to apply the periodic BC, can I use your " 0 = N-1, N+1 = 2 " for all the three grids INDEPENDENTLY? Or, there has to be some correlation between applying periodicity for the three variables (may be because of staggered grid).

Since my results are still diverging, I can not understand where I am implementing BC wrong (my code works for Dirichlet BC where I specify a fix value on all the boundaries).

[sbaffini]

You should be able to use it independently for each grid. Maybe is something hardcoded in the implicit part?

[kaya]

Originally Posted by tanmayagrawal7

Thanks for your answer!

So, as in my case of a staggered grid, I have three separate grids for U, V and P. Their grid sizes being [N,N+1],[N+1,N],[N+1,N+1] respectively.

For the U velocity, I solve NS equations for the interior points i.e. upto N-1 points in x and N points in y-direction, and similarly for the other two primitive variables. I am using an explicit finite difference based method.

So, to apply the periodic BC, can I use your " 0 = N-1, N+1 = 2 " for all the three grids INDEPENDENTLY? Or, there has to be some correlation between applying periodicity for the three variables (may be because of staggered grid).

Since my results are still diverging, I can not understand where I am implementing BC wrong (my code works for Dirichlet BC where I specify a fix value on all the boundaries).

For the sake of simplicity lets say your indices start from 1 and N=5 i=1,2,3,4,5 j=1,2,3,4,5




P(N,j)=P(1,j) j=1...N
P(i,N)=P(i,1) i=1...N

u(N,j)=u(1,j)
v(i,N)=v(i,1)

[sbaffini]

I might have seriously overlooked at the matter (i only did staggered grid once, around 10 years ago, and without periodicity). Actually, what i suggested works properly for colocated finite difference schemes.

Now i took my old program and put the variables down, and it is more clear to me.

When you go to staggered you need to do different things for different variables/direction. In practice the grid has to be considered a la finite volume, which means you have:

1) Nx x Ny cell centers, where pressure is stored
2) (Nx+1) x Ny x-normal faces where Ux velocity is stored
3) Nx x (Ny+1) y-normal faces where Uy velocity is stored

Thus, what to do? For the pressure, you strictly work a la finite volume:

0x = Nx
0y = Ny
Nx+1 = 1x
Ny+1 = 1y

For the Ux velocity you work a la finite volume along y, but finite difference along x:

0x = Nx-1
Nx+1 = 2x
0y = Ny
Ny+1 = 1y

For the Uy velocity it is the converse, finite difference along y and finite volume along x:

0y = Ny-1
Ny+1 = 2y
0x = Nx
Nx+1 = 1x

This implies that your variables are decleared as:

P(Nx,Ny), Ux(Nx+1,Ny), Uy (Nx,Ny+1)

This should work but, as i said, i never tried.

[ened89]

@sbaffini For an arbitrary problem and staggered scheme; lets say we have another variable (e.g shear stress component) grid Sxy (Nx+1) x (Ny+1) what has to be done?

[ened89]

For an arbitrary problem and staggered scheme; lets say we have another variable (e.g shear stress component) grid Sxy (Nx+1) x (Ny+1) what has to be done?

Originally Posted by sbaffini

I might have seriously overlooked at the matter (i only did staggered grid once, around 10 years ago, and without periodicity). Actually, what i suggested works properly for colocated finite difference schemes.

Now i took my old program and put the variables down, and it is more clear to me.

When you go to staggered you need to do different things for different variables/direction. In practice the grid has to be considered a la finite volume, which means you have:

1) Nx x Ny cell centers, where pressure is stored
2) (Nx+1) x Ny x-normal faces where Ux velocity is stored
3) Nx x (Ny+1) y-normal faces where Uy velocity is stored

Thus, what to do? For the pressure, you strictly work a la finite volume:

0x = Nx
0y = Ny
Nx+1 = 1x
Ny+1 = 1y

For the Ux velocity you work a la finite volume along y, but finite difference along x:

0x = Nx-1
Nx+1 = 2x
0y = Ny
Ny+1 = 1y

For the Uy velocity it is the converse, finite difference along y and finite volume along x:

0y = Ny-1
Ny+1 = 2y
0x = Nx
Nx+1 = 1x

This implies that your variables are decleared as:

P(Nx,Ny), Ux(Nx+1,Ny), Uy (Nx,Ny+1)

This should work but, as i said, i never tried.

[FMDenaro]

Originally Posted by ened89

For an arbitrary problem and staggered scheme; lets say we have another variable (e.g shear stress component) grid Sxy (Nx+1) x (Ny+1) what has to be done?

This is not a new variable, you compute it from the velocity.

[sbaffini]

Originally Posted by ened89

For an arbitrary problem and staggered scheme; lets say we have another variable (e.g shear stress component) grid Sxy (Nx+1) x (Ny+1) what has to be done?

What is exactly your question? What has to be done for what?

[ened89]

I have a scheme as in the picture for an FDTD simulation. I want to use periodic BC in left and right, for velocity components I do the same procedure but I dont know what to do for Sxy to make periodicity. Without periodicity I am making stress mirroring and the become zero in boundary.

Originally Posted by sbaffini

What is exactly your question? What has to be done for what?

[ened89]

So should I compute again Sxy from the copied velocity values ?

Originally Posted by FMDenaro

This is not a new variable, you compute it from the velocity.

[sbaffini]

Very generally speaking, for periodicity (but applies to the distributed parallel case as well) one should only exchange the independent variables (say, velocity) and compute everything in place. So, as mentioned by Filippo, the stress should be recomputed once you have the correct velocity in the periodic nodes.

EDIT: ok, this doesn't sound really correct in general, but for staggered grids might be your best option

[sbaffini]

Originally Posted by ened89

So should I compute again Sxy from the copied velocity values ?

yes indeed