Thanks for the detailed answer, Marcos. There surely seems to be a strong dependance of the method on k-sampling used. Unfortunatelly it also reveals some general weaknesses of SIESTA in application to complex structures. SIESTA is a great piece of DFT sofware, that could handle systems built of thousands atoms, had'n it had a disastrous electronic minimisation algorithm in scf. To take an obvious example, none of my 'big' systems (>500 atoms) has ever converged at a higher k-sampling if I started from DM generated at gamma point. There are not too many ways of setting kpoins for such systems. The electronic loop always 'runs out of control' even at MixingWeight as low as 0.01. Of course, geometry of the structures was optimized beforehand up to 0.04 eV/atom. (In fact, I would really appreciate if somebody from the community could provide me with an example of a structure that really benefits from the reusing DM generated at a lower kpoint set) And SIESTA finds groud state when I restart the simulations from scratch, though it may take forever until it converges. Sadly, when number of atoms is less than 700, even 'heavy' VASP (special gamma point version) outperforms SIESTA, I understand because of its robust and reliable algorithms set for scf as well as reusing WAVECAR for restarts that helps a lot. Still, the work that was done by the developers cannot be underestimated, and what I especially appreciate is making SIESTA free to all!
Thanks for reading, Max 22 октября 2011, 16:21 от Marcos Veríssimo Alves <[email protected]>: Maxim, I believe Chun has a point here, and I beg to differ from what you say. For some materials, it is important to include special k-points, so as to get a correct band structure and (eventually) forces - the most "famous" case would perhaps be graphene, with its K point. Even if your material hasn't such special characteristics, making the k-point mesh denser means that you are sampling the BZ more accurately and thus changing the part of the total energy that corresponds to the electronic band occupations. Just think of a simple numerical integration in 1-D. To say that you have a problem with an increasing k-point mesh density is a bit precipitated, unless you are sure that you have a very converged k-mesh. Chun's e-mail is actually very sensible: you should first relax your structure, and then increase the k-mesh to see its effects on the forces.Not only that, you should also check its effects on the band structure, although in principle this effect should be negligible, if you have performed a convergence study. Marcos On Sat, Oct 22, 2011 at 6:34 AM, Maxim Peskov <[email protected]> wrote: Thanks for the comment, Chun. However, you have missed the point. I guess it will be better if I rephrase my question: how can one restart from scf-preconverged density matrix and increased kpoint set? With the options I used such a restart would mean the forces are going crazy which clearly indicate a problem. Even more, if I change kpoints it is better to start from scratch than to use DM. Any ideas? --Max 20 октября 2011, 20:04 от [email protected]: > Hi, > > From the first ourput file, I found the following lines: > > outcoor: Final (unrelaxed) atomic coordinates (fractional): > -0.00005348 0.49985936 0.49676159 1 1 C > ... > ... > > which means your first run is not converged. > > Best, > > Chun > ---------------- С уважением, Maksim V. Peskov
