Top Page > Browsing
MD step update and mixing weight; specification or problem?
Date: 2017/08/01 18:10
Name: Asako Terasawa

Dear those who concerned,

I found the following issue about the update of mixing weight in standard output log.

The mixing-related parameters in the input file is as follows.

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

As intended, scf calculation started from initial mixing weight:

******************* MD= 1 SCF= 1 *******************
<DFT> Mixing_weight= 0.002000000000

Then, scf calculation in MD = 1 ended at a larger mixing weight:

******************* MD= 1 SCF=114 *******************
<DFT> Mixing_weight= 0.030000000000

However, when updated, the mixing weight starts from the maximum weight:

******************* MD= 2 SCF= 1 *******************
<DFT> Mixing_weight= 0.300000000000

It is understandable that this may cause just a minor effect if a single MD update causes a very small change in electronic states, which may not be the case sometimes.
So, I want to know that if this behavior is intended or not.

Page: [1]

Re: MD step update and mixing weight; specification or problem? ( No.1 )
Date: 2017/08/01 23:11
Name: T. Ozaki


During MD, the initial charge density at each MD step is obtained by extrapolation using
charge densities at previous steps. So, in many cases the estimated charge density
can be a good approximation for the converged one, and we expect that convergence can be
obtained even using a large mixing weight. That's why Mixing_weight is set to the maximum

However, as you worry, such a treatment may cause charge sloshing problem again for some
cases. For such cases it might be better to use a relatively smaller scf.Max.Mixing.Weight.

If you do not want such a behavior, you only have to comment out the following line
of 905 in DFT.c:

if (MD_iter!=1) Mixing_weight = Max_Mixing_weight;

If you do that, Mixing_weight is not set to Max_Mixing_weight.


Re: MD step update and mixing weight; specification or problem? ( No.2 )
Date: 2017/08/02 12:33
Name: Asako Terasawa


Thank you for your very helpful answer.
Now I and one of colleague are trying both suggestions and checking the behavior of NormRD and mixing_weight, because we are treating systems which easily diverge.
I will let you know again if I get any suggestive results.

Re: MD step update and mixing weight; specification or problem? ( No.3 )
Date: 2017/08/31 13:49
Name: Asako Terasawa


After your suggestion, I examined the behavior of convergence with changing the mixing weight at MD update. The detail of the calculation is presented in the pdf:
(I will delete it sooner or later)

It seems that the changing of the mixing weight doesn't have an effect on the instability of SCF convergence and so far I don't know the reason for such a behavior, whether it comes from the mixing, the extrapolation or the calculated system itself.
But I want to note here is that, there is a case that SCF convergence at the previous time step does not necessarily mean that the convergence at the next step is stable.
Therefore, for future release (or some other options), we wish to have some alternative option for determination of mixing parameters at MD update.

Or alternatively, any suggestion will be appreciated that indicates the inappropriateness of the parameter setting or reasonableness of the behavior of NormRD, or anything else.

Best regards,
Re: MD step update and mixing weight; specification or problem? ( No.4 )
Date: 2017/08/31 14:04
Name: Asako Terasawa

Sorry, I forgot to write a somewhat important information in the pdf.

In each figure, graphs are aligned in the following order:
(1) (2)
(3) (4)
(5) (6)
and this means that the NormRD convergence is shown for each time step in order.
Re: MD step update and mixing weight; specification or problem? ( No.5 )
Date: 2017/09/07 17:01
Name: T. Ozaki


As long as I understood from your data, the instability appears for cases
with electric field. Is your system an isolated or slab model?
If your system is bulk, the use of scf.Electric.Field is not justified theoretically.
Even if the system is an isolated or slab model, the unit cell has to be large enough
so that the discontinuity in the sawtooth potential can be far from the system.
Otherwise the charge will be trapped at the dip in the sawtooth potential, which may
cause such an instability in the SCF process.



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