Personally I would only trust an 100 Ry cutoff calculation to provide a starting point for a geometric configuration destined for further relaxation. My experience tells me if you use oxygen you should always be using 350 Ry or more, and typically for some materials I've been simulating recently, as high as 550Ry. Incidentally, depending on your simulation cell, any cutoff will set the *actual* grid used to be higher, sometimes much higher. If you take a stable geometry and perform a single-point energy (0 CG steps) convergence test with respect to the meshcutoff, try and see what happens when you increment the mesh cutoff by steps of about 100 Ry or so all the way to about 2000 Ry- you should find that the energy won't change by much after a certain point, and the cutoff you should choose will balance being respectably close to that converged energy. Always record the meshcutoff used by siesta, not the one in your input file.
For example, a while ago with bulk TiO2, I found that the difference in total energy between a ~1400 Ry cutoff and a ~550 Ry cutoff was about the same as the difference between ~450 Ry and ~550 Ry. I chose the 550 Ry grid as it was 'sufficiently converged' for my purposes, did an egg-box test, printed the proof in a couple quick plots and saved it somewhere accessible. Having a tight grid that is balanced by its memory usage is a trick you kind of have to get a knack for. It is true that some problems really do require tight expensive grids and a lot of memory, and it might seem to squeeze a lot of the time benefits out of using LCAO-based approach (vdW xc-functionals come to mind). In these cases your parallelization scheme and how you select initial reference geometries will make a lot of difference! The "egg-box" convergence test is another one you will find yourself doing often to check the reliability of your numerical grid, look elsewhere on the list for that. The bottom line is, I would always explicitly include the MeshCutoff as an input in an fdf of mine for research purposes. Even when I need a 'quick and dirty' estimate of something or perhaps a stepping stone to a reference geometry to iron out some of the more expensive force-fluctuation math. On Wed, Apr 11, 2012 at 6:36 PM, Yi Gao <[email protected]> wrote: > > Dear siesta users, > > Recently I have been using siesta to find optimized structure of crystal > such as anatase. > > According to what I have found, if we run the CG or Broyden method to > search for atom positions with least forces, say, 0.02 eV/Ang, the default > MeshCutoff (e.g., 100.0 Ry) gives a configuration. However, if I simply > displace the origin of my lattice from somewhere to some atom, keeping the > primitive vectors unchanged, the forces exerted on each atom will change, > which is unphysical. > > What I can do is to increase the MeshCutoff to try to get converged > results, i.e., displacing the origin of the lattice does not change the > forces on each atom, and finally in agreement with experiments. But it > turns out to be MeshCutoff = 500.0 Ry, so large that it is difficult to > extend to larger systems. > > My question is that how can I trust the atom configuration with the > default MeshCutoff? Is it always necessary to set MeshCutoff to an > exceedingly large value to get converged results? > > Best wishes! > > Yi > > -- Abraham Hmiel Katherine Belz Groves Fellow in Nanoscience Xue Group, College of Nanoscale Science and Engineering at SUNY Albany
