RMM Convergence Issue with LEPSILON=.TRUE.

[michael_wolloch]

Dear Sari,

I am happy to hear that.

Just to confirm:
Setting

Code: Select all

LMAXMIX = 6

did the trick? Could you isolate the contribution of setting this tag?

If you have no further questions, I will mark the issue as resolved and lock the thread.

Cheers, Michael

[lisg_sari]

Dear Michael,

For this case LMAXMIX = 6 or LMAXMIX = 4 does not make difference as I just tested.
AMIX and BMIX settings somewhat change the convergence.
I initially tried:

AMIX = 0.01
BMIX = 0.0001
AMIX_MAG = 0.8
BMIX_MAG = 0.0001

The response in the 2nd axis did not converge well, (still close to the converged energy).
Then I switched back to the default, which I used for optimizations, it worked better.

Most likely, the new version 6.4.3 has some new features for linear response calculations, which are helpful for this system.

[michael_wolloch]

Dear Sari,

I apologize for not checking earlier, but indeed there were some improvements to the stability of linear response calculations in 6.4.3. The default for CSHIFT was changed, along with other tweaks to make linear response calculations more robust. These changes are responsible for the improved convergence you are seeing and are mentioned in the changelog on the wiki. Note that in the current version of the code, for linear response theory with DFPT, CSHIFT is hardcoded to 0.02, and setting CSHIFT in the INCAR will not change the internal value, although the set value will be reported in the OUTCAR in the linear response block. This is unfortunate. I created an issue and we will adapt the code and/or the documentation.

Thanks for getting back to me again and also testing the other mixing. Can I lock this topic now?
Michael

[lisg_sari]

Sure.
Thanks a lot for clarifications for the improvement and really helpful.
Hopefully it will be helpful for other systems too.