How many empty bands should be used in GW calculations for large supercell ?

Submitted by dean on Tue, 07/05/2016 - 02:53

I am calculating a big supercell system which contains about 1000 electrons, do I have to include more than 10 times empty bands in epsilon and sigma calculations? The computing resources consumption is huge.

Submitted by dyq on Wed, 07/06/2016 - 10:49

For every system, you should always test the convergence of your GW calculation with respect to the number of empty bands in epsilon and sigma. Otherwise, you have no way of determining how accurate/meaningful your results are. Testing the convergence with bands can be done with a smaller number of k-points to save computational cost.

If the computational cost of the final production run is too high, some possible options you can try are: 1) run with a smaller number of bands and accept and report a larger error, 2) try to come up with a reasonable way of extrapolating your results from a smaller system or less converged result, 3) try an implementation of GW without empty states. There is some recent work on GW formalisms without empty states using the Sternheimer approach. This is not currently implemented in BerkeleyGW, and I'm not really familiar with the codes that use this, so I don't know at all if it will be faster for your case, but you might give it a try. One such code is GWL.