Open pohlan opened 3 years ago
Experimenting with the numerical parameters showed that dτ_h
(pseudo-time step for h
) cannot be chosen completely randomly. If it is too large, ϕ
has problems to converge, and if it is too small h
and also ϕ
converge very slowly. Furthermore it seems to depend on the grid resolution.
I tried making it somehow proportional to Res_h
, Res_ϕ
or dτ_ϕ
but couldn't figure out anything useful. I only managed to get the expected model output for a constant dτ_h
determined through try and error.
dh/dt
is the recation-part of the \phi
equation. Is there something about how to choose a time step for a reaction-diffusion equation?
But h
is not just a reaction of ϕ
diffusion, it influences the diffusion itself (with the factor h^α
in the effective diffusivity). Possibly that's a reason why the h
pseudo-time step is limited somehow. In this specific case, the h
values mostly increase over pseudo-time, so if h
advances quickly the effective diffusivity is very large, such that ϕ
can only take tiny steps and doesn't converge well. This would fit together with the observation that ϕ
generally converges better for small dτ_h
, but then h
itself converges very slowly and if dτ_h
is too small the error just produces NaN
s. Anyway, this is just some guessing, the issue might as well be somewhere else.
But what I don't quite understand is why it doesn't work well if ϕ
and h
have the same pseudo-time step.
oo small the error just produces NaNs
Why is that?
dh/dt = vo-vc
and vo-vc
features as reaction term. Hmm, but you're right, it probably does not matter.
For the diffusion part of
ϕ
, we use the CFL criterion, what about the other terms includingvo
,vc
,m
? This would especially be relevant forh
, there we only use a randomly determined scalar for all grid points at the moment.What should be the relation to
dt
, the physical time step? min(CFL, dt), a harmonic mean or something else?