Top Page > Browsing
Convergence problem?
Date: 2017/04/24 11:24
Name: Asako Terasawa

Dear those who concerned,

I post this issue to report or ask about a seemingly strange convergence problem in an OpenMX calculation.

I have a difficulty in getting a converged result in my calculation.
To figure out a reason, I checked out the NormRD value in the standard output.
And I found some strange behavior as follows.

The following table shows the extracted MD, SCF and NormRD values from the standard output.

110 0.000338002901
111 0.000108841524
112 0.000050245356
113 0.000069157162
114 0.000081580317
115 0.000046530677
116 1.000000000000 <-
117 0.639339774899
118 4.225276759412
119 3.149034771858
120 1.400572326300
290 0.000420898438
291 0.000142576793
292 0.000362888509
293 0.000297455627
294 0.000089358157
295 0.000592933083
296 1.000000000000 <-
297 1.037622154842
298 2.857130788009
299 1.344543027898
300 0.724770578156
520 0.000212969000
521 0.000114067661
522 0.000104224630
523 0.000340883890
524 0.000093475218
525 0.000102990222
526 0.000299032848
527 0.000284832542
528 1.000000000000 <-
529 0.765202061879
530 5.833908520858

This means that, NormRD suddenly goes back to 1 when the system seems to reach convergence from the previous steps.
This happens in every few hundred steps for this time step.

I use OpenMX 3.8, compiled in issp sekirei by myself with patches.
And calculation options related to SCF is as follows:

scf.SpinPolarization On # On|Off
scf.ElectronicTemperature 3000.0 # default=300 (K)
scf.energycutoff 600.0 # default=150 (Ry)
#scf.Ngrid 32 32 32
scf.maxIter 2000 # default=40
scf.EigenvalueSolver band # Recursion|Cluster|Band
scf.Kgrid 7 7 7 # means n1 x n2 x n3
scf.Mixing.Type rmm-diisk # Simple|Rmm-Diis|Gr-Pulay
scf.Init.Mixing.Weight 0.0020 # default=0.30
scf.Min.Mixing.Weight 0.0001 # default=0.001
scf.Max.Mixing.Weight 0.3000 # default=0.40
scf.Mixing.History 30 # default=5
scf.Mixing.StartPulay 10 # default=6
scf.criterion 1.0e-6 # default=1.0e-6 (Hartree)

Currently I could not find the straightforward reason for this, so I would like you to help me for this problem.
I would appreciate if you give me any idea or suggestion for this problem.

Asako Terasawa
Page: [1]

Re: Convergence problem? ( No.1 )
Date: 2017/04/24 12:40
Name: Kylin

My openmx configuration for a Mg twin crystal at T=700 K

scf.Mixing.History 60 # default=5
scf.Mixing.StartPulay 40 # default=6
scf.Mixing.EveryPulay 1 # default=5

Just try to increase increase the number of step in dat file.

However for the jump of NormRD back to 1, I didn't known or review the openmx code.
But I guess openMX may think the mixture procedure was located on some local minima. Thus the reset back to 1 was another trial to find the convergent result.

Honestly, the charge mixture and the convergence among the SCF loops were a greate issue in DFT calculation, you should try to increase the electronic thermal temperature and the mixing configuration for your system.

Re: Convergence problem? ( No.2 )
Date: 2017/04/24 13:57
Name: Asako Terasawa

Thank you for a quick response!
I submitted a calculation with following your instruction, and waiting for the job to run.
I will let you know in the case that the same problem happens.

Asako Terasawa
Re: Convergence problem? ( No.3 )
Date: 2017/04/27 11:30
Name: Asako Terasawa


I'm sorry, this now seems to be my miscoding of the extraction script of NormRD.
I didn't recognize that the "MD= 9" was separated by space, where as "MD=10" is not in the standard output.
Therefore my script could not find where is the change of MD step.

Considering this point, my calculation seems to converge faster by using the parameter you suggested.

Many thanks,
Re: Convergence problem? ( No.4 )
Date: 2017/05/18 11:05
Name: T. Ozaki


During the SCF calculations, a temporal file, *.DFTSCF, is geneated.
The file might be useful to check the SCF convergence on-the-fly,
while the file is integrated to *.out, and deleted after finishing
the calculation.


Re: Convergence problem? ( No.5 )
Date: 2017/08/01 18:07
Name: Asako Terasawa


Thank you so much, for I've just found your reply.
This is pretty much easier to look at.


Page: [1]

Title (must) Move the thread to the top
Your Name (must)
E-Mail (must)
Password (used in modification of the submitted text)
Comment (must)

   Save Cookie