I think at this point I may be ready to just say that indeed IBI cannot be used to converge to a potential that is able to reproduce the structure of xtalline materials (or at least the material I am studying).
I've tried
(1) diminishing the factor used to update the potential (as you mentioned) and it did not work.
(2) updating literally only one potential at a time in the IBI and keeping the others literally constant either in the BI potential or in analytical forms that are able to reproduce perfectly the probability distributions. This would discard the possibility of dependence on the degrees of freedom in that sense that the update of one potential is affecting the distributions related to other potentials.
(3) Although the result is not meant to be bin-size-dependent, I tried playing with the bin size of both, the references I am feeding to VOTCA, and of the distributions it is meant to built as the iterative process runs for the different potentials. I thought maybe I was not setting up "proper" bin sizes for the algorithm.
(4) I tried dividing the angles lying within each of the two peaks in the initial figure I showed into two different angle types and it also did not work.
(5) I read your paper and tried to be more careful with issues that you raised in section 2.9 related to the smoothness of the distributions in the onset region (although VOTCA is supposed to take care of this internally apparently via the extrapolation methodology). Although section 2.10 bring up issues related to IMC, I also tried some more ideas that came to mind from reading that section and it didnt work.
(6) I've tried keeping analytical forms for the bonded potentials (I happen to have analytical forms that perfectly reproduce the distributions) and optimize the non-bonded and it also doesnt work.
Naturaly, in all cases, together with weird distributions, my potentials are also going to hell as the iterative procedure goes on (which explains why the corresponding distributions are weird).
For sure the problem doesnt have to do with the "sharpness" of the probability distribution curves (due to the xtalline material being highly ordered) cause I tried to feed "artificial" target distributions that are wide and thus less step and I dont converge to anything reasonable either.
Maybe the shape of the distributions for xtalline materials is not friendly to be used within IBI to converge to a potential, idk...
Well..