I agree with what you have mentioned. But as far as I understand, when I use a smaller mixing parameter, the convergence will be smooth but need more steps. However, the fact is not always the case: when the iteration is very near the convergence but hard to converge, I use a much smaller mixing parameter to restart the task, it will not improve the convergence.
I agree, not least about the observation that there are no 100% rules here. Also note that the history steps can almost have a larger influence than the mixing parameter.
Also, the convergence behaves very well for some bias (especially the low bias) and when the bias increase especially there are transmission peak entering the bias window, the convergence will be very difficult. However, for some system, even if there are transmission peak entering the bias window, there is no convergence problem. What does it imply for the former case? Does the corresponding state of the tranmission peak not couple to the electrodes well at that energy?
Since you probably climb in bias (using the 0.1 V converged state as starting guess for 0.2, etc), it seems reasonable that if the 0.2 case has a resonance peak and 0.1 doesn't, the converged 0.1-state is not, after all, a very good starting guess for 0.2. There are no shortcuts in this game