There is not much to do other than trial an error. Typically one can get a better convergence by increasing the number of layers that separate the junction and the electrodes. I.e. if the electrostatics are long-range. Also, when you have vacuum regions as terminations of the buffer layers you could also increase the number of layers "behind" the electrode to better screen the surface.
Other than that it is difficult, it is purely trial and error. You could try with siesta-4.1-b2 which has better convergence, and is _much_ faster than the previous versions of transiesta. 4.1 also enables one to fix the DM in the electrodes to match that of the pristine electrode (see flag: TS.Elecs.DM.Init). This _may_ improve a bit the convergence. 2017-06-20 14:31 GMT+02:00 Leonardo Fonseca <[email protected]>: > Dear all, I have been converging transiesta 4.0 without problems for a > number of systems, including graphene and gold chains. However, I recently > tried to calculate the transmission function for Ti/HfO2/TiN and could not > make it converge. The DM part converges fine, but the GF part oscillates. I > tried longer buffer regions, more k-points and smaller mixing parameters > (0.0002 right now), to no avail. Then I replaced TiN by Ti to simplify the > problem. With that change the size of the system decreased from 258 to 120, > by eliminating vacuum along z and and the necessary buffer layers. The GF > part also oscillates. Finally, I also noticed that upon restarting the job > in the GF part in case it has not converged, the progress obtained in the > previous job is totally lost, and major oscillations in the total energy > take place for several steps. Any ideas on how to improve the convergence > in situations like this? > -- Kind regards Nick
