I'm having some trouble getting this ferromagnetic state to converge with DFT+U. The input file below has U=6eV. It converged fine for U=0,2,4eV with Broyden mixing and a reducebf=0.75.
After the 6eV one failed with the above settings I tried it with linear adaptive mixing and upped the reducebf to 0.85 as below, but I still can't get it to converge. This also failed for U=8,10eV.
I'm not sure what other things I can tweak to try and get the convergence right... so any help is much appreciated!
you do not explain how it fails. Does it converge to zero moment or does it not at all converge?
Without knowing I have a problem to help. If it is non-conergence you might have to lower the beta0 parameter.
However I can say that it usually better to start from an already converged solution if you increase a parameter as U, that is use task 1 instead of 0. In addition I think it is odd to use J=0 as in your calculation. Are you sure that is what you want? There is also other parameters that I personally would not have chosen. I have my clear ideas how to calculate magnetic solutions for metallic actinide compounds and how to interpret the possible appearance of small moment solutions. You can find them published under first author Cricchio and Bultmark if you are interested.
Good luck,
Lars
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks for the detailed response, I'd like to address all the points you made...
Sorry for not being clear on the convergence. It reaches the maximum number of self consistent loops. At this point the energy is converged to the target precision but the KS potential is not. So i suppose your suggestion would be to further decrease beta0 then?
For using task 1 instead of 0, do you mean to use a ground state density obtained with U=0? Or a lower value of U for which it has converged? And what are the parameters that you would not have chosen?
My understanding of the Elk implementation of DFT+U is that there is an effective Ueff=U-J parameter (which i think I read somewhere deep on this forum). Based on what you said it seems this is not the case...
Thanks for the reference, I am in fact hoping to get moments of around 1/3 so this could be useful.
Thanks,
Harry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
if its not converge in the maximum steps you might increase
maxscl
600
or you can restart you calculation
by rerunning it a second time with
setting
tasks
1
at this stage you might also reduce the initial magnetic fields global or local to 1D-4. The suggestion of lars in this respect was to use you convergered computation for U=4 and then set U=6 in the input and switch to task 1, so you are already starting from a kind of very converged density. In this resepect you might increase
gmaxvr
18
or
gmaxvr
20
and add
trimvg
.true.
which helps to reach convergence in some cases. Last, please note that the implementation of U and J is not simply U-J, as in done in other codes pseudopotential codes, but instead a full treatment of both values.
best
Michael
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks! I like the idea of using task 1 to go through U values so I'll give that a go.
And since my understanding of DFT+U in Elk is clearly flawed, are you aware of a good reference for the implementation of DFT+U in Elk?
Thanks,
Harry
Last edit: Harry K 2017-07-05
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi All,
I'm having some trouble getting this ferromagnetic state to converge with DFT+U. The input file below has U=6eV. It converged fine for U=0,2,4eV with Broyden mixing and a reducebf=0.75.
After the 6eV one failed with the above settings I tried it with linear adaptive mixing and upped the reducebf to 0.85 as below, but I still can't get it to converge. This also failed for U=8,10eV.
I'm not sure what other things I can tweak to try and get the convergence right... so any help is much appreciated!
Regards,
Harry
tasks 0 xctype 20 dft+u 1 1 1 3 0.2204958 0 autolinengy true mixtype 1 beta0 0.01 rgkmax 7.0 gmaxvr 15 spinorb true maxscl 200 nempty 20 sppath '/work/y07/y07/cse/elk/elk-4.0.15/species/' reducebf 0.85 avec 8.987000000 0.000000000 0.000000000 -4.493500000 -7.783011610 0.000000000 0.000000000 0.000000000 5.880000000 atoms 2 : nspecies 'U.in' : spfname 1 : natoms; atpos, bfcmt below 0.00000000 0.00000000 0.00000000 0.00000000 0.00000000 1.00000000 'Au.in' : spfname 2 : natoms; atpos, bfcmt below 0.3333333333 0.6666666666 0.50000000 0.00000000 0.00000000 0.00000000 0.6666666666 0.3333333333 0.50000000 0.00000000 0.00000000 0.00000000 ngridk 6 6 9Dear Harry,
you do not explain how it fails. Does it converge to zero moment or does it not at all converge?
Without knowing I have a problem to help. If it is non-conergence you might have to lower the beta0 parameter.
However I can say that it usually better to start from an already converged solution if you increase a parameter as U, that is use task 1 instead of 0. In addition I think it is odd to use J=0 as in your calculation. Are you sure that is what you want? There is also other parameters that I personally would not have chosen. I have my clear ideas how to calculate magnetic solutions for metallic actinide compounds and how to interpret the possible appearance of small moment solutions. You can find them published under first author Cricchio and Bultmark if you are interested.
Good luck,
Lars
Hi Lars,
Thanks for the detailed response, I'd like to address all the points you made...
Sorry for not being clear on the convergence. It reaches the maximum number of self consistent loops. At this point the energy is converged to the target precision but the KS potential is not. So i suppose your suggestion would be to further decrease beta0 then?
For using task 1 instead of 0, do you mean to use a ground state density obtained with U=0? Or a lower value of U for which it has converged? And what are the parameters that you would not have chosen?
My understanding of the Elk implementation of DFT+U is that there is an effective Ueff=U-J parameter (which i think I read somewhere deep on this forum). Based on what you said it seems this is not the case...
Thanks for the reference, I am in fact hoping to get moments of around 1/3 so this could be useful.
Thanks,
Harry
Dear Harry,
if its not converge in the maximum steps you might increase
maxscl
600
or you can restart you calculation
by rerunning it a second time with
setting
tasks
1
at this stage you might also reduce the initial magnetic fields global or local to 1D-4. The suggestion of lars in this respect was to use you convergered computation for U=4 and then set U=6 in the input and switch to task 1, so you are already starting from a kind of very converged density. In this resepect you might increase
gmaxvr
18
or
gmaxvr
20
and add
trimvg
.true.
which helps to reach convergence in some cases. Last, please note that the implementation of U and J is not simply U-J, as in done in other codes pseudopotential codes, but instead a full treatment of both values.
best
Michael
Hi Michael,
Thanks! I like the idea of using task 1 to go through U values so I'll give that a go.
And since my understanding of DFT+U in Elk is clearly flawed, are you aware of a good reference for the implementation of DFT+U in Elk?
Thanks,
Harry
Last edit: Harry K 2017-07-05
OK, I would follow Michael's suggestions.
As for references, I suggest the ones I already gave.
Good luck,
Lars