27 Aug
2019
27 Aug
'19
10:29 p.m.
On 27/08/2019 16:45, Shawn W Walker wrote:
Just a small adjustment to this.
If you fix the pressure at a node, then solve, you should then subtract off the mean value of the soln so it is mean value zero. Even if you fix it at a corner (like he suggests) there can be a slight degradation in the convergence (in L^2) by a log factor. I know you can compute the mean value fairly easily in FEniCS; I assume you can do it in FireDrake as well.
Great point. You should also manually orthogonalise if you specify the nullspace to PETSc, as PETSc enforces orthogonality in the Euclidean inner product: https://github.com/firedrakeproject/firedrake/issues/1329 Patrick