[MITgcm-devel] Re: [MITgcm-cvs] MITgcm/pkg/mom_vecinv CVS Commit

[Alistair Adcroft]

I've checked in an alternative calculation of length scale which I hope 
will be more robust across different types of grid; it takes into 
account the aspect ratio of grids but is not so precise that it works 
for strong local variations (this would involve a lot of i+1,j-1's).

It is turned off be default; I've introduced an in s/r logical that 
needs to be turned on: useSophisticatedLengthScale=.TRUE.

The new length scale is used only in the maximum viscosity calculations 
so shouldn't change anything that isn't limited. I did originally want 
to change Leith and grid-dependent formula but since it's already in 
such wide-spread use I think it's simpler to leave that all unchanged.

In principle, for a smooth grid, forward time-stepping and no other 
explicit terms, a value of viscA*GridMax=1.0 should be stable; in 
practice it needs to be reduced by about 1/4. I tested it for an 
anisotropic grid as follows:

  set viscA4Grid=100 (very large on purpose)
  set viscA4GridMax=1.0, run the model which blows up quickly
  reduced to viscA4GridMax=0.5, run the model which blows up
  etc..
  found that viscA4GRidMax=0.3 ran for quite a while
  found that viscA4GridMax=0.25 is robust.

I'm sure that a factor of 1/3 is due to Adams-Bashforth II and the 
remaining is due to other explicit terms.

Give it a try, let me know if it doesn't work as expected.

If it does work then it might be worth using the same length scale in 
the grid-dependent viscosity because it will mean that coefficients of 
order 1 will be meaningful.

A.

-- 
Dr Alistair Adcroft              email: Alistair.Adcroft at noaa.gov
NOAA/GFDL, Princeton University               Tel: (609) 987-5073
Forrestal Campus, U.S. Route 1, P.O. Box 308, Princeton, NJ 08542