PAW+phonons

[ksmirnov]

Hello everybody.

Trying to test the phonon calculations with LDA + PAW I've encountered the following problem. Everything went well as soon as the calculations dealt with a small size cell system (such as Li2O). However, trying to test the calculation of phonons at Gamma for alpha-quartz and alpha-cristobalite, I obtained three large negative frequencies (-83 cm-1) for the former structure and four (!) negative frequencies (-162.7, -130.1, -125.2, -90.4 cm-1) for the latter. The situation becomes worse for a-V2O5, where a third of phonon frequencies are negative. Of course, the complete structure optimization is done before performing the phonon calculations. PAW files were taken from ATOMPAW database; small radius LDA file was used for oxygen to avoid overlapping of PAW spheres.

Did anybody notice similar problem. Thank you.

Konstantin.

[ebousquet]

Dear Konstantin,

First of all, did you check the convergence with ecut with these Atomic data on your systems?
If yes, did you check if the SCF for each perturbation is converged up to the tolerance you required in your input (when looking at "ETOT" in the log or ouput files for each perturbation)?

All the best,
Eric

[ksmirnov]

Dear Eric,

Since I did the calculations to test the capability of ABINIT, I did not test the convergence with ecut. The ecut used was 25 Ha with pawecutdg 50 Ha. The information in the atompaw database shows that such cut-off should be enough. By the way, increasing ecut to 30 Ha does not change the results.

I inspected the log and output files. The SCF converges after ca. 30 steps for each perturbation (for alpha-cristobalite), the convergence criterion tolvrs=1.00E-15.

Best regards,

Konstantin.

[ilukacevic]

Dear Eric,

Since I did the calculations to test the capability of ABINIT, I did not test the convergence with ecut. The ecut used was 25 Ha with pawecutdg 50 Ha. The information in the atompaw database shows that such cut-off should be enough. By the way, increasing ecut to 30 Ha does not change the results.

I inspected the log and output files. The SCF converges after ca. 30 steps for each perturbation (for alpha-cristobalite), the convergence criterion tolvrs=1.00E-15.

Best regards,

Konstantin.

Hi, Konstantin!

Also, did you check the convergence wrt other parameters, like ngkpt?

And also did you use the same parameters for the relaxation and for the phonon calculations?

Best regards,

Igor

[ksmirnov]

Hello Igor,

Also, did you check the convergence wrt other parameters, like ngkpt?

And also did you use the same parameters for the relaxation and for the phonon calculations?

I use kprtlatt grid suggested by ABINIT, the convergence parameters are identical to those used in the geometry optimization. Pseudopotential calculations (*.fhi TM NCPP) using EXACTLY the same criteria give correct results. I 've just tested ecut=40 Ha (pawecutdg is always taken as 2 x ecut) for alpha-cristobalite -> the same weird results, i.e. four large negative frequencies.

Best wihes,
Konstantin.

[suvas]

Dear Igor,

I have a question for code developers. Is response function is implemented for PAW in abinit 7.1.2 version. I have been waiting PAW implementation for phonon calculation for long. I can calculate phonon frequency using norm conserving plane wave pseudopotential without problem. But if I use PAW pseudopotential, my calculation crahses when it reaches to 3DTSET ie., perturbation with repsect to homogeneous electric field. I checked this even in latest version abinit 7.2.1., the problem remains same. This time error message is different than in abinit version 6.12.2.

I would be so grateful if you could reply me soon.


Regards,

Suvas

[jzwanzig]

can you post your input file?

[suvas]

Thanks for your reply. Here is my input file attached.

[ksmirnov]

can you post your input file?

I suppose that the request was for me. The input file for the phonon calculations is attached. The ecut is enormous (40 Ha) and was used just to test the convergence issue. The atomic PAW files are:

Oxygen - http://users.wfu.edu/natalie/papers/pwp ... index.html (LDA, Rc=1 Bohr)
Silicon - http://users.wfu.edu/natalie/papers/pwp ... index.html (LDA)

Best regards,

Konstantin.

acrb-phonons.in

(2.83 KiB) Downloaded 458 times

[ebousquet]

I suppose that the request was for me. The input file for the phonon calculations is attached. The ecut is enormous (40 Ha) and was used just to test the convergence issue. The atomic PAW files are:

Oxygen - http://users.wfu.edu/natalie/papers/pwp ... index.html (LDA, Rc=1 Bohr)
Silicon - http://users.wfu.edu/natalie/papers/pwp ... index.html (LDA)

Dear Konstantin,

The atomic data coming from N. Holzwarth web site are known to be ecut consuming...
I propose you to use the one from GPAW code:
https://wiki.fysik.dtu.dk/gpaw/setups/setups.html

Or the one coming from Rabe/Vanderbilt group (only PBE up to know, so DFPT is not yet working for GGA):
http://www.physics.rutgers.edu/gbrv/

With these atomic data, you should get convergence with much smaller ecut than 40 Ha

Check if the phonons look better, for the moment I don't see what is wrong.
Please use the latest version of Abinit (7.2.2).

Best,
Eric

[ebousquet]

The atomic data coming from N. Holzwarth web site are known to be ecut consuming...
I propose you to use the one from GPAW code:
https://wiki.fysik.dtu.dk/gpaw/setups/setups.html

Or the one coming from Rabe/Vanderbilt group (only PBE up to know, so DFPT is not yet working for GGA):
http://www.physics.rutgers.edu/gbrv/

With these atomic data, you should get convergence with much smaller ecut than 40 Ha

By the way, on the abinit web site if you have recommanded atomic data availble, they are also very good.
The one from contributed users can also be OK, Joe can you confirm for the one you generated?

All the best,
Eric

[ksmirnov]

The atomic data coming from N. Holzwarth web site are known to be ecut consuming...
I propose you to use the one from GPAW code:
https://wiki.fysik.dtu.dk/gpaw/setups/setups.html

Or the one coming from Rabe/Vanderbilt group (only PBE up to know, so DFPT is not yet working for GGA):
http://www.physics.rutgers.edu/gbrv/

With these atomic data, you should get convergence with much smaller ecut than 40 Ha

Please use the latest version of Abinit (7.2.2).

Dear Eric,

The ecut of 40 Ha was just for testing as the first calculation I did with ecut=30 Ha.

I'll test other PAW files (in particular those you recommended) as soon as I can and shall post the results.

The calculations were done with the version 7.2.1

Best regards,

Konstantin.

[ksmirnov]

These are the results for alpha-cristobalite with other PAW files. Parameters: ecut=30 Ha, geometry optimization with
tolmxf=1.0d-6, SCF criterion tolvrs=1.0d-15 is identical in both the geometry optimization and response calculations.

Phonon frequencies in cm-1 :
- -4.024624E+02 -3.493627E+02 -3.338475E+02 -2.761988E+02 -2.032222E+02
- 2.762025E+01 2.815139E+01 1.416090E+02 1.776531E+02 1.814254E+02
- 2.217504E+02 2.449746E+02 2.975136E+02 3.068896E+02 3.545199E+02
- 3.741691E+02 4.268524E+02 4.329080E+02 4.821977E+02 4.841514E+02
- 4.888155E+02 6.927731E+02 7.322794E+02 7.993585E+02 8.209918E+02
- 8.345867E+02 8.455979E+02 8.748048E+02 1.030342E+03 1.043978E+03
- 1.044429E+03 1.051239E+03 1.089727E+03 1.111666E+03 1.177921E+03

Five large negative frequencies!

The PAW files are :
http://www.abinit.org/downloads/PAW2/TA ... -hard-uspp - recommended by ABINIT site.
http://www.abinit.org/downloads/PAW2/TA ... i-ca--vgrp - compatible with the PAW file for oxygen, ixc=2.

Any ideas?

Konstantin.

[ebousquet]

Yes, these phonons sounds quite odd...
Like that I cannot tell you what can be wrong, do you mind to put your output file in case I can see any trouble?
All hte best,
Eric

PS: btw, did you try to run with 7.2.2 version instead of 7.2.1?

[ksmirnov]

Yes, these phonons sounds quite odd...
Like that I cannot tell you what can be wrong, do you mind to put your output file in case I can see any trouble?

PS: btw, did you try to run with 7.2.2 version instead of 7.2.1?

Dear Eric,

A set of input/output files is attached. What is a bit troubling to me is that when I start in the phonon run from the optimized geometry, the GS calculation yields the forces on atoms more than by an order of magnitude larger than those at the end of the optimization run (cf the two entries below) :

frms,max,avg= 1.7126812E-07 2.1720023E-07 0.000E+00 0.000E+00 0.000E+00 h/b
frms,max,avg= 3.7619149E-05 7.9186500E-05 0.000E+00 0.000E+00 0.000E+00 h/b

I did not try the 7.2.2 version yet, shall do that ASAP.

Best regards,
Konstantin

P.S. The attached file is actually in tgz format. I had to change the file extension to get the file attached.

[ksmirnov]

Results from 7.2.2 version are:

Phonon frequencies in cm-1 :
- -4.024624E+02 -3.493627E+02 -3.338475E+02 -2.761988E+02 -2.032222E+02
- 2.762020E+01 2.815136E+01 1.416090E+02 1.776531E+02 1.814254E+02

The same weird behaviour.

Nice WE to everybody.
Konstantin

[ebousquet]

Dear Konstantin,

To know if this coming from pseudos, code or whatever, I did quickly a frozen phonon calculation (with phonopy) with your input files and your psp's and I don't have any instability:

- q-position: [ 0.0000000, 0.0000000, 0.0000000 ]
band:
- # 1
frequency: -0.9597835076
- # 2
frequency: -0.9597835075
- # 3
frequency: -0.3448233495
- # 4
frequency: 45.0782282831
- # 5
frequency: 121.7301843603
- # 6
frequency: 155.5141475474
- # 7
frequency: 155.5141475474
- # 8
frequency: 229.0974909380
......etc.......
- # 36
frequency: 1169.6592734575

So this means that there is a problem somewhere with the DFPT (code, input file, compatibility with uspp, ...) that I do not understand.
Let me forward the problem to the PAW-DFPT person...
Best regards,
Eric

[ksmirnov]

To know if this coming from pseudos, code or whatever, I did quickly a frozen phonon calculation (with phonopy) with your input files and your psp's and I don't have any instability:

So this means that there is a problem somewhere with the DFPT (code, input file, compatibility with uspp, ...) that I do not understand.
Let me forward the problem to the PAW-DFPT person...

Dear Eric,

First of all, excuse me for a silence, was out of lab for few days. I also think that there is a problem with the DFPT part, more specifically at the point dealing with PAW (DFPT+NCPPs works fine). As I noticed in my first post, the problem appears while moving to systems with a (relatively) large number of atoms in the cell (for alpha-V2O5 I got a half of frequencies negative). Thank you for the time and efforts. Now, wait and see...

Best regards,

Konstantin

[Boris]

Hi Konstantin,

The phonon frequencies calculated at Gamma with PAW+DFPT are wrong because of the breaking of the acoustic sum rule. This is related to the fact that the system should be invariant by translation and it is actually not the case. The phonon frequencies therefore need to be corrected and this can be done with the anaddb code (see tutorial) by setting ASR = 1.

Some other hints: always use the latest version of abinit because the PAW+DFPT part still changes and set USEXCNHAT to 1. Also, phonon frequencies are strongly dependent on the quality of the ground-state wavefunctions, therefore you should converge your wavefunctions to at least TOLWFR = 1.0D-22. Convergence of the wavefunctions can be improved by increasing NLINE (I set it to 20)

Good luck
Boris

[ksmirnov]

Dear Boris,

Thank you for the post.

The phonon frequencies calculated at Gamma with PAW+DFPT are wrong because of the breaking of the acoustic sum rule. This is related to the fact that the system should be invariant by translation and it is actually not the case. The phonon frequencies therefore need to be corrected and this can be done with the anaddb code (see tutorial) by setting ASR = 1.

The analysis of phonons was done with the anaddb code and ASR was set to 1.

Some other hints: always use the latest version of abinit because the PAW+DFPT part still changes and set USEXCNHAT to 1. Also, phonon frequencies are strongly dependent on the quality of the ground-state wavefunctions, therefore you should converge your wavefunctions to at least TOLWFR = 1.0D-22. Convergence of the wavefunctions can be improved by increasing NLINE (I set it to 20)

Following the input variable page of ABINIT, I used the tolvrs as the convergence criterion (set to 1.d-15 in both the geometry optimization and response calculations). Do you know a relation between the tolvrs and tolwfr? I shall test you suggestions and keep you informed.

Best wishes,

Konstantin

[Boris]

I don't know the relation between tolvrs and tolwfr, but you can have tolvrs converged to 1.d-15 and still have wavefunctions not converged enough for a DFPT calculation.

I would use tolwfr=1.d-23 for the ground state calculation, then tolvrs=1.d-8 for the subsequent RF calculations.

Boris

[ksmirnov]

Dear Boris,

Some other hints: always use the latest version of abinit because the PAW+DFPT part still changes and set USEXCNHAT to 1. Also, phonon frequencies are strongly dependent on the quality of the ground-state wavefunctions, therefore you should converge your wavefunctions to at least TOLWFR = 1.0D-22. Convergence of the wavefunctions can be improved by increasing NLINE (I set it to 20)
Boris

I tested your suggestions and neither resulted in correct behaviour of frequencies. ABINIT 7.2.2, the USEXCNHAT parameter was set to 1 (by the way it is the default value), TOLWFR = 1.0d-22 in the GS calculation, and TOLVRS=1.d-15 is the response one. The result is

Phonon frequencies in cm-1 :
- -4.024624E+02 -3.493627E+02 -3.338475E+02 -2.761988E+02 -2.032222E+02
- 2.762021E+01 2.815136E+01 1.416090E+02 1.776531E+02 1.814254E+02

that is the same as before.

Konstantin.

[Boris]

Thanks for the feedback.

Could you please try with ecut = 25 Ha and pawecutdg = 40 Ha?

Thanks
Boris

[ksmirnov]

Could you please try with ecut = 25 Ha and pawecutdg = 40 Ha?

Results are:

Phonon frequencies in cm-1 :
- -4.067343E+02 -3.536478E+02 -3.402707E+02 -2.818307E+02 -2.100758E+02
- 4.004261E+01 4.529726E+01 1.446989E+02 1.820860E+02 1.861747E+02

One negative frequency less, but what is really the reason of taking pawecutdg < 2.*ecut. Is there any educated guess?

Konstantin.

[Boris]

pawecutdg doesn't necessarily need to be 2*ecut. It just has to be greater than ecut. It should be set as a convergence criterium, like the kpt mesh or the cutoff energy.

As for your results, something is troubling me. You said that you analyzed the frequencies with anaddb, but anaddb is supposed to impose the acoustic sum rule, therefore you should have the first three frequencies set to zero. Plus, what you are quoting in your posts look like the abinit results printed in the output file, are you sure you're looking at the output of anaddb?

Boris