Numerical instability with split_collisions?
Activating split_collisions removes the collision operator from the response calculation and changes the time advance algorithm to first take a full collisionless step, then apply collisions and finally recalculate the fields to complete the step.
At low ky and in simulations including A|| one can find a numerical instability/blowup when the collision frequency is suitably large. The value of vnewk to trigger the instability appears to depend on the species mass.
The instability is not there if only the ediffuse operator is applied. It is not there if const_v = T. PR #614 does not appear to help improve this. Dropping the timestep doesn’t appear to help, suggesting it may not be cfl based.
Whilst the instability appears to require sufficiently large vnewk, it goes away if one ensures the collision frequency entering the lorentz matrix (~vnewk/energy^1.5) doesn’t drop below a certain value – i.e. boosting the higher energy collision frequencies.
It is also possible to push the non-split collisions unstable if vnewk is sufficiently large. It’s not clear if this is the same form of instability as seen with split collisions or not.
