[petsc-users] Preconditioner for stokes flow with mixed boundary conditions.

Mon Apr 19 23:42:22 CDT 2021

  You might consider a discretization that replaces the direct 4th order discretization with a coupled set of second order discretions and then using PCFIELDSPLIT to organize appropriate algebraic multigrid on the 2nd order discretions.  The thing is traditional linear iterative solvers are totally focused on 2nd order systems and if you blindly give them a 4th order system they will choke 99% of the time. PCFIELDSPLIT is our attempt to allow one to split up a system and use appropriate preconditioners on the parts but the parts likely need to be second order for the traditional preconditioners to work.

  Direct solvers don't care about the order, but don't scale. There are a few specialized 4th order iterative solvers but it is generally easier to make multiple 2nd order systems and use the traditional iterative solvers.

  Barry

> On Apr 19, 2021, at 1:55 PM, Abhinav Singh <abhinavrajendra at gmail.com> wrote:
> 
> The only reference that I know of which solves these equations is :
> https://www.pks.mpg.de/fileadmin/user_upload/MPIPKS/group_pages/BiologicalPhysics/juelicher/publications/2015/AHp-MMfIAPVG2015.pdf <https://www.pks.mpg.de/fileadmin/user_upload/MPIPKS/group_pages/BiologicalPhysics/juelicher/publications/2015/AHp-MMfIAPVG2015.pdf>
> 
> There is coordinate free form in the appendix. They have been solved using the UMFPACK Solver on staggered grids. I am trying a new approach (pressure correction with auxiliary potential). In 2 dimensions, the approach worked well with GMRES. In 3d, GMRES again works but with Dirichlet boundary conditions.
> 
> On Mon, 19 Apr 2021 at 20:45, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> wrote:
> On Mon, Apr 19, 2021 at 2:38 PM Abhinav Singh <abhinavrajendra at gmail.com <mailto:abhinavrajendra at gmail.com>> wrote:
> Hello,
> 
> The Stokes flow equations are a 3d version of the equations attached (Stokes-Leslie Flow). Only variables/unknowns are v and u_xy=0.5(Dx(vy)+Dy(vx).
> 
> This is not what I have referred to as Stokes flow (Google gave me no results for Stokes-Leslie flow). The Stokes operator is elliptic, but I have no idea if
> what you have written is. It has fourth order derivatives in it, with mixed nonlinear terms, so nothing is clear to me. Is there a coordinate-independent form?
> From what I see in the image, I have no idea what solvers might work. Do you have any reference where people have solved this before?
> 
>   Thanks,
> 
>      Matt
>  
> I am trying to solve them iteratively by correcting the pressure to reach a steady state. I start with 0 pressure. Currently, I am unable to solve the first iteration (periodic in X and Z. V=0 at Y=0 and Dx(vy)+Dy(vx)=g(x) at Y=10).
> 
> I think the equations might be singular but I am not sure as in my experience, the problem is well posed if the solution is known at certain boundaries.
> 
> 
> <by default 2021-04-19 at 20.28.28.png>
> 
> On Mon, 19 Apr 2021 at 15:59, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> wrote:
> On Mon, Apr 19, 2021 at 9:37 AM Abhinav Singh <abhinavrajendra at gmail.com <mailto:abhinavrajendra at gmail.com>> wrote:
> What does this mean? Stokes means using an incompressibility constraint, for which we often introduce a pressure.
> Yes, what I mean is solving the momentum block, say with known pressure. viscosity is constant however, the momentum equation has both Laplacian and Gradient terms.
> 
> That does not make any sense to me. Can you write the equation?
> 
>   Thanks,
> 
>      Matt
>  
> You should use a good Laplacian preconditioner, like -pc_type gamg or -pc_type ml.
> I tried gamg and it seems to diverge as the solution is NaN. The KSP residual message is " 0 KSP Residual norm 1.131782747169e+01  ".
> When using -pc_type ml, I get Aggregate Warning and then some faulty address which stops the code.
> 
> 
> 
> On Mon, 19 Apr 2021 at 12:31, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> wrote:
> On Mon, Apr 19, 2021 at 6:18 AM Abhinav Singh <abhinavrajendra at gmail.com <mailto:abhinavrajendra at gmail.com>> wrote:
> Hello all,
> 
> I am trying to solve for incompressible stokes flow on a particle based discretization. I use a pressure correction technique along with Particle strength exchange like operators.  
> 
> I call Petsc to solve the Stokes Equation without the pressure term.
> 
> What does this mean? Stokes means using an incompressibility constraint, for which we often introduce a pressure.
> 
> Do you mean you are solving only the momentum block? If so, do you have a constant viscosity? If so, then this is just the Laplace equation.
> You should use a good Laplacian preconditioner, like -pc_type gamg or -pc_type ml.
> 
>   Thanks
> 
>      Matt
>  
> GMRES usually works great but with dirichlet boundary conditions. When I use a mixed boundary condition in Y, (dirichlet on bottom and Neumann on the top) with periodicity in X,Z. GMRES fails converge when the size of matrix increases. For smaller size (upto 27*27*5), only GMRES works and that too only with the option 'pc_type none'. I was unable to find any preconditioner which worked. Eventually, it also fails for bigger size. UMFPACK works but LU decomposition fails after a certain size and is very slow.
> 
> It would be great if you could suggest a way or a preconditioner which suits this problem.
> 
> Kind regards,
> Abhinav 
> 
> 
> -- 
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener
> 
> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
> 
> 
> -- 
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener
> 
> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
> 
> 
> -- 
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener
> 
> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210419/8e6d860a/attachment.html>