This fixes a long-standing problem in the computation of the RHS of the potential equation in Grid_SolveForPotential.C. This was pointed out by Dan Reynolds last year, but I didn't work out the ramification and the fix until now. The basic problem is that Grid_SolveForPotential.C assumes the cell spacing on the grid is L/Nx, while the multigrid solver assumes it is L/(Nx-1). As Dan points out, this means that we end up solving a problem of the form
Laplace(Potential) = (G/a)(GravitatingMass)(dim1-1)(dim2-1)(dim3-1)/dim1/dim2/dim3
as opposed to:
Laplace(Potential) = (G/a)*(GravitatingMass)

I spent some time figuring out to change the multigrid solver, but that turned out to be a bad idea, since it works best with that cell spacing. Instead, it is much easier to adjust the RHS computation. Note these changes are completely separate from the recent discussion about depositing subgrid cells.

This will require a regeneration of the gold standard.

Comments (3)

Sam Skillman

Hi Greg,

This seems like an automatic accept to me. Any reason why it shouldn't be accepted, and have the gold standard regenerated?

Greg Bryanauthor

Yes, I think it should be accepted. Thanks for the scripts!

Tom Abel

Fantastic sleuthing, Greg! It also gives a noticeable improvement at the grid boundaries. I'll try my TotalJeansRefinement ideas again now. This could have been what was missing. Thanks!

Hi Greg,

This seems like an automatic accept to me. Any reason why it shouldn't be accepted, and have the gold standard regenerated?

Yes, I think it should be accepted. Thanks for the scripts!

Fantastic sleuthing, Greg! It also gives a noticeable improvement at the grid boundaries. I'll try my TotalJeansRefinement ideas again now. This could have been what was missing. Thanks!