It appears that previous attempts at filtering out population inversion forgot to set some levels to LTE, due to only doing a single pass over all populations. To make sure this is fixed, the following heavy-handed approach has been used:
do two passes (respectively starting from the lowest and the highest level) and afterwards set all affected levels simulatenously to LTE
repeat this the until no extra levels need to be changed
Note to self: also check if Ng-acceleration could cause some additional issues.
Edit: Ng-acceleration seems to have no bad influence on the convergence
It appears that previous attempts at filtering out population inversion forgot to set some levels to LTE, due to only doing a single pass over all populations. To make sure this is fixed, the following heavy-handed approach has been used: