How to converge natural convection with bad mesh.
 

I'm simulating with Fluent (Icepak) with a very skewed mesh, and we cannot simplify geometry at this point.

Transient natural convection with Ra ~ 1e8. Been using laminar solver in an older version of Icepak/Fluent, which detects divergence but corrects itself.

In old Fluent/Icepak, warning is:

# Divergence detected in AMG for pressure correction: protective actions enabled!
# Divergence detected in AMG for pressure correction, temporarily solve with BCGSTAB!

This happens frequently during iterations, but simulation still converges/completes.

With new Fluent/Icepak, however, we get this:

Stabilizing pressure correction to enhance linear solver robustness.
Stabilizing pressure correction using GMRES to enhance linear solver robustness.
Divergence detected in AMG for pressure correction
Error at host: floating point exception

It seems like new Fluent/Icepak does not have the same protective mechanisms as old Icepak.

1. How to stabilize natural convection with a bad mesh?

2. How to use whatever old Fluent/Icepak was using to achieve a complete solution?