Convergence study

Moderators: gmatteo, bruneval, rigna

Post Reply
Posts: 16
Joined: Tue Jan 19, 2010 2:31 pm
Location: Zurich, Switzerland

Convergence study

Post by Martin » Sat Jun 12, 2010 8:57 pm


I am planning to perform a convergence study for a t-ZrO2 unit cell and several defect supercells (up to 96 atoms). Acc. to the first ABINIT tutorial on GW the following steps are to be taken [parameter(s)]:

1. Fully converged LDA/ or PAW ground state calculation [ecut, pawecutdg], incl. k-point convergence [ngkpt, nshiftk, shiftk]
2. Calculation of the Kohn-Sham structure (KSS file); k-point convergence [ngkpt, nshiftk, shiftk]
3. Convergence on the number of planewaves in the wavefunctions to calculate the Self-Energy [ecutwfn3].
4. Convergence on the number of planewaves to calculate Sigma_x [ecutsigx].
5. Convergence on the number of bands to calculate the Self-Energy [nband3].
6. Convergence on the number of planewaves in the wavefunctions to calculate the screening (epsilon^-1) [ecutwfn4].
7. Convergence on the number of bands to calculate the screening [nband4].
8. Convergence on the dimension of the epsilon^-1 matrix [ecuteps].

The physical quantity of interest is the band gap and (when feasible) the full band structure. I am using a python script to do the convergence job.

I still have some questions on the topic:

Question A: Can the convergence of the GS calculation be treated separately, i.e. with respect to the LDA/ PAW band gab?
Question B: When can the calculation be considered converged? Is it a meaningful criterion that the absolute difference in band gap energy is below a certain limit, e.g. 10meV (twice, successively)?
Question C: I only have access to limited computational resources (for now ~16 cores, SGI Altix). Which limit would you suggest (1meV, 10meV, or even 100meV)?
Question D: How can I estimate the accuracy of the final result? Something like plus/minus 100meV?
Question E: What about k-point convergence? Same convergence criterion?
Question F: Of course, the overall accuracy of the result depends on how well the chosen method (in this case PAW/LDA + one shot GW) applies for the given physical system. Therefore the chosen pseudopotential (LDA) or the PAW atomic data (PAW) is a very important factor! How can I evaluate these, when no experimental data is available for the defect supercells? Is it a "good guess" to use the same pseudopotential /atomic data as for the unit cell? What if the defect significantly changes the electron density? Do I have to think about this (plotting out electron density, compare to chosen core radii)?


Best regards,

Martin Haeufel,
TU Munich, WSI (T33)
(5.74 KiB) Downloaded 200 times

Post Reply