Large DOS calculations

[elj]

I'm currently calculating the DOS for several large metal/adsorbate systems. The largest of these, however, is simply too big a calculation to undertake at once. I've been experimenting with trying to split the run up and treating the k-points individually (or as small groups), as described in the manual. I've done this by using the IBZKPT file from a previous run.

However, the manual recommends that one use the Tetrahedron Method (ISMEAR = -5) for DOS calculations. To enter the k-points explicitly using this method you need the tetrahedra connection list. I cannot find how to calculate this when I am extracting just some of the k-points from the previous run.

How, therefore, can I perform this calculation? Is it okay not to use the tetrahedron method in this case, or is there another way of generating the connection list. Or can I only treat one k-point at a time? (annoying, as there is a considerably number of k-points in this system).

Many Thanks

[cpp6f]

Large systems don't need very many k-points. The number of k-points needed for a given accuracy scales inversely with the system size. For a 3x3 surface unit cell, a 4x4x1 Monkherst-Pack k-point mesh is sufficient (however this is with ISMEAR=2; ISMEAR=-5 will probably require a coarser mesh for the same accuracy). This mesh samples 48 points in the Brillouin zone (24 irreducible points - this is the actual number of points that are calculated). This mesh uses an 8x8x1 grid to divide up the Brillouin zone. An odd mesh size such as a 3x3x1 would use a 3x3x1 grid to divide up the Brillouin zone - even meshes use twice as many divisions as odd meshes along a given reciprocal space lattice vector. For a 6x6 unit cell you would want a 4x4x1 division of the grid, so use a 2x2x1 mesh. For a 9x9 unit cell you would want an 2.67x2.67x1 ~ 3x3x1 division of the Brillouin zone, so use a 3x3x1 mesh. Since the number of k-points scales inversely with system size and memory requirement at a fixed number of k-points scales quadratically with system size, the required memory will scale linearly with system size to achieve a given level of accuracy - until you go past a 24x24 unit cell at which point you will be using a 1x1x1 k-point mesh that cannot be reduced any further.

[tjf]

[quote="'smallblacktext'>[ Edited Fri Aug 25 2006, 01:22PM "]</span>

[cpp6f]

I was referring to the system that elj descibed above (metal/adsorbate). I have found this mesh to give binding energies that converge to within ~0.01 eV for many late transition metal/adsorbate (containing C,H,and O) systems. I meant to say only that it's a good starting point.