About rescaling k-point grid in supercell calculations

[alpinnovianus]

Dear all,

I have a question about rescaling k-point grid in supercell calculations.
I read that in some papers when a calculation is done in a supercell (say twice or quadruple the size of primitive cell), the author rescaled down the k-point grid by half or a quarter to "maintain the same k-point grid density". I know this is related to the BZ getting smaller in size when the computational cell gets larger.
But, what I want to know is, how important is maintaining the grid density to our calculations?

For example, if we use the same k-point grid in a supercell that is twice as large as the primitive cell, what is the problem?
By using same k-grid in unit and supercells, aren't we just sampling the BZ more finely in the supercell? Please let me know if there is a known pitfall.

My opinion is that grid rescaling is a good-to-have adjustment for comparison between unit cell and supercell results, but it is not that critical: it doesn't invalidate the results of the supercell calculations as a standalone calculation. Am I mistaken ?

If rescaling the k-point grid is a must-do, would this mean we have a difficulty to compute say, the density-of-states in a reasonably large supercell?
Because DOS calculation requires a fine mesh, rescaling the grid would mean we may be left with a countably coarse mesh (like a grid fewer than 4x4x4, or even only the Gamma point) when we have to use large supercells.

Thank you for your sharing.

[ferenc_karsai]

For example, if we use the same k-point grid in a supercell that is twice as large as the primitive cell, what is the problem?
By using same k-grid in unit and supercells, aren't we just sampling the BZ more finely in the supercell? Please let me know if there is a known pitfall.

Yes you are sampling more finely, but at the same time it costs much more computational resources. If a calculation gets computationally very demanding you probably want minimize to the least required number of k points that show satisfactory convergence. For that you would first do the convergence tests on the smaller cells and then run with the same k point distance in the larger cell.

My opinion is that grid rescaling is a good-to-have adjustment for comparison between unit cell and supercell results, but it is not that critical: it doesn't invalidate the results of the supercell calculations as a standalone calculation. Am I mistaken ?

The whole purpose as described above is to minimize computational effort.