Sorry, I forgot another important parameter. Try higher values of SCF.ElectronicTemperature, this will also improve your convergence (this really helps a lot, at the expense of results quality). Try using the Methfessel-Paxton smearing function. I think that a second-order MP polynomial should be good.
2006/12/5, Vasilii Artyukhov <[EMAIL PROTECTED]>:
Dear SuiYang: > Your mixing parameter is too large > redata: New DM Mixing Weight = 0.2500 > that's why (at least) you get no convergence. > Try to set > DM.MixingWeight .05 > (or less), to begin with. > There are two important issues to consider: > 1) in a large supercell, there is genererally more difficult to achieve > convergence because you have many bands with contributions from similar > atoms around the Fermi level. So the system has a tendency > to swap charge from one atom to another, which effect you might need to > damp > more eficiently than was the case in a smaller cell. In the single cell > with only few bands and few atoms, you won't normally have this > "degree of freedom" in your system, so the convergency goes > sraightforwardly. > 2) In a metal system, depending on the actual band sructure, > the accuracy of the k-summation/interation becomes an important issue > influencing the accuracy of your result, and the convergence to it. > In SIESTA where no tetrahedra integration is available, you must > carefully test the effect of increasing number of k-points, and > the energy broadening used in the search for the Fermi level. > Your present choice 5*5*5 k-points with ElectronicTemperature 300 K > is not a priori unreasonable for getting a preliminary convergence, > but be careful to make additional tests if you want to extract > some energy differences afterwards. A comment on that: the problem you are facing is called "charge sloshing", you might want to look it up in literature. As Dr. Postnikov says, dince the electrons near the Fermi level are strongly delocalized, when solving the SCF equations iteratively, you get oscillations of the electron density with long wavelengths. This problem can be overcome in a number of ways. For instance, one could use a multi-level real-space approach when the SCF equations are first solved on a coarse grid to eliminate the long-wavelength contribution, then corrected on a finer grid, and the process is repeated until the convergence is reached (but that's not something you can do within the LCAO approach). Another way is to use Kerker density (or potential, whatever) mixing when the mixing is done in the Fourier space with low-frequency component damping. This scheme is available in the OpenMX package, but unfortunately, it is not implemented is SIESTA. So, it seems that the best you can do in SIESTA is decrease the mixing weight until you get some convergence (DM.MixingWeight). You might also want to try different values of mixing history, DM.NumberPulay (also pay attention to the SCF.NumberKick and DM.KickMixingWeight parameters, they might help). Finally, you might want to skip the Pulay mixing at all and do simple linear mixing with an appropriate mixing weight. This could require a number of restarts (in order to use different mixing weights at different levels of convergence: for instance, high, then low, then high again). Don't forget to use the DM.MixSCF1 variable when restarting. Finally, a word regarding the k-point grids: although a dense k-point grid is essential for an adequate description of properties of metals, I don't think that denser grids could help you improve your SCF convergence, since the SCF equations are solved at each k-point independently (which is why the k-point parallelization is so efficient). Therefore, I suggest thet you try and find the parameters that give you convergence with a sparse grid, and then transfer them to denser grids. Regards, Vasilii
