EPW Forum

Dear all users,

Recently, I have ran the GaN example and successfully obtained the phonon linewidth by setting the parameter phselfen and a2f to be true. However, when I draw the phonon dispersion curve, I find that the phonon self_energy of the 12nd optical branch is very large, to be 10^4. I want to know what causes that problem. The detailed parameters I use for calculations are like this:

For phonons,
k-mesh 6x6x6
q-mesh 6x6x6

For epw,
k-mesh 50x50x50
q path choose the gan_band.qpt

Is it the problem of unconvergence?

Best regards,
Jiayue Yang

RWTH Aachen University

Dear Jiayue,

You are probably obtaining this behavior because the matrix elements for the highest optical branch diverge as 1/q (Frohlich coupling) as described for example in the new EPW paper http://arxiv.org/pdf/1604.03525v1.pdf
The calculation of phonon linewidths for a semiconductor is a bit tricky because you need to have a well defined Fermi level (see Eq. (4) in the paper), i.e. you need to dope the material which brings different physics to the simple Frohlich coupling.

Best
Carla

Dear Carla,

Thanks for your advice and I have read the linked paper. It does a great favor. Yet I have another problem on how to simulate the doping in EPW. Should we just assume that the doping only affects the Fermi energy, or should use the virtual crystal approximation to build the doped pseudopotential?

Best
Jiayue

RWTH Aachen University

Dear Jiayue,

You can use:

1) rigid band approximation: you just shift the Fermi level and assume that this will not affect the shape of the bands.

2) virtual crystal approximation: you create a psp with fractional occupation. This assumes that the doping is very homogeneous.

3) "real doping" : You do supercells calculations and you replace some of the atoms by their doping counterpart. Off course you have to make a lot of configuration since
the doped element can be in a lot of different configurations (for example cluster of doping elements might be possible).

Expected accuracy of the methods: 3 > 2 > 1
Expected computational cost of the methods: 3 >>> 2 >= 1

I hope this helps,

Samuel