Nicola Marzari wrote: > > > Nektarios, > > > 1) the "dispersion" from the magnetic impurities should be flat, > and the difference in energy that you look at is within two > cells that have the same sampling. That is good. > > 2) a simple test would be to still use 1 k-point, but the Baldereschi > (PRB 1973) point instead of Gamma. For an orthorombic cell, I generally > use 1/4 1/4 1/4 . PWSCf would automatically try to replicate it to > recover full symmetry (i.e. would want to use all of the 2 2 2 1 1 1 > Monhorst-Pack mesh that are not related by any point-group operation), > but using no_sym (or nosym ?) you can force the code to just use that > single k-point. 1/4 1/4 1/4 is often more accurate than Gamma, and > any difference you find would hint at insufficient sampling.
Thanks, that is a great idea! > > 3) degauss, generally speaking, should be larger than what you used if > you were dealing with a metal. In your case, though, you probably use it > just to smooth the path to selfconsistency, and at the ground state you > still want integer occupations everywhere, correct ? At the very end (after ion relaxation also) I would like degauss=0. I guess if the variation has lead to a semiconducting structure that should be OK. If the structure has a metalic character then the oscillation effects will not allow for convergence with degaus=0. But I havent tried sofar. > > 4) what I would worry is haivng very good ionic relaxations, since these > can be different in the FM and AFM states. Using the same cell for the > two calculations (same cell parameters, same cutoff, same kpoints) helps > a lot with the electronic accuracy, since in the energy difference you > have cancellation of terms that might have needed more kpoints. I am thinking to allow for relaxation of the first neighbohrs of the deffects (impurities / native deffects). > > > 5) of course, if you need for some reason absolute k-point convergence, > than you need to increase the number of k-points, but as long as you > compare unit cells that are identical, you might easily get away with > lower sampling. Still, gamma can be slow to converge - have a look > at the literature on the vacancy formation energy in silicon. > > nicola > Thank you very much for the advice Nektarios > > >> Hi all, >> >> I want to calculate the energy difference of FM and AFM states of two >> isolated >> magnetic impurities in a semiconductor, optionaly with the presence of >> native >> impurities. >> >> The supercell has to be fairly large, of the order of 100 atoms, so I >> can only >> afford Gamma-point calculation. The difference in the energy of >> different magnetic states might be of the order of mRyd. >> >> My question concerns the convergence parameters and especially degauss. >> Right now I use degauss=5.d-4 and thats the smallest I can use without >> serious oscillation problems. That is 0.5mRyd so it is probably fine. >> >> A second concern is the realiability of the Gamma point calculation >> itself. >> My worry is that increasing the mumber of k-points will lower all the >> energies >> by much more than the mRyd differences I am obtaining. In other words, >> I am worrying that the error of restricting the k-space to the gamma >> point >> is much larger than the energy differences I get. What is the >> experience on that >> issue? Is a supercell of 100 atoms big enough to guarantee that a >> gamma-point >> calculation will have accuracy of the order of miliRydberg? And if not >> are the >> energy differences of any value or it is just crap? >> >> Here are the convergence parameters I use right now: >> etot_conv_thr = 1.d-5 >> degauss = 5.d-4 >> conv_thr = 1.0d-6 >> >> Thanks >> Nektarios >> _______________________________________________ >> Pw_forum mailing list >> Pw_forum at pwscf.org >> http://www.democritos.it/mailman/listinfo/pw_forum > > >
