Dear Kay,
I understood that we can break spin symmetry for magnetic calculation with a small bfcmt or bfieldc.
Which routine should I take a look to understand the symmetry breaking by these fields ? I'm trying to understand the optimal ratio between bfcmt and bfieldc to study the extermal magnetic field effect.
Best,
bbbusybee
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Take a look at the routine 'seceqnsv'. The second-variational Hamiltonian is of the form
H = H_0 + sigma.B + H_so
where H_0 is the first-variational Hamiltonian (diagonal in the first-variational states) and H_so is spin-orbit coupling. The sigma.B term is the one your interested in: both bfcmt and bfieldc are applied like this in exactly the same way. The only difference is that bfcmt is in the muffin-tin, and bfieldc is applied everywhere.
The ratio shouldn't matter. Most of the time, the B-fields are used to break spin symmetry in a particular way.
Cheers,
Kay.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Using an AMF double-counting with Slater parameters, this converges, but very slowly (about 1,5 days with 40 cores).
I wanted to try using U + automatically determined Yukawa screening length and an interpolation between AMF and FLL (i.e. dftu 3, inpdftu 5), however the calculation takes even longer (more than 2 days with the same setup).
I was thinking of ways to speed this up:
Should I add reducebf 0.5? If so, which should I increase, bfieldc 2.0 2.0 0.0 or bfcmt to -2.0 -2.0 0.0 (or both)? Will this give significant improvement in convergence speed?
I read that Broyden mixing "only works if a large field is provided" in the Fe-fast example. Does that mean my current setup gives inaccurate results?
Will there be any significant difference, especially in band structure and RPA results, between the AMF and the AMF-FLL Interpolation?
I'm a bit confused since the parameters are very dependent of one another.
Best,
Cenna
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
your input is a little bit counteracting. You are applying a local magnetic field on Cu to enforce a AFM state and at the same time you also apply a global magnetic field, which favors a FM state. From my experience I would suggest you to do the followin:
1) remove bfieldc
2) increase the field at the atoms to a value between 1 - 5
3) set reducebf = 0.8 -- 0.9
4) set mixing to linear
this should speed up things a little bit. Actually, broyden mixing (for me in transitionmetal oxides) is not always really faster.
Regarding your third question
--> "Will there be any significant difference, especially in band structure and RPA results, between the AMF and the AMF-FLL Interpolation?"
This depends in my experience Cuprates tend to be more FLL. I would suggest here the following strategy first compute FLL and AMF and than you can run the mixing scheme (or all three at the same time if your cluster allows). There is a file written out (alpha.OUT) or something where elk tells you, for the FLL-AMF mixed scheme, what the actual interpolation parameter is. If it is away from 0 (meaning actually AMF) than you may can expect differences.
hope this clarifies
best regards
Michael
PS: It would be nice to post your observation here, what the actual alpha in the interpolation scheme is and if it makes a differnence.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thank you so much for you input. Currently testing your suggestions and I'll share with you when it's done.
Is there any particular reason for using linear mixing instead of the default adaptive mixing?
I've done AMF and the FLL-AMF interpolation so far, here's the alpha values
1 1 3 : species, atom, l
0.3655505842E-02 : alpha
1 2 3 : species, atom, l
0.3655505842E-02 : alpha
1 3 3 : species, atom, l
0.3678230245E-02 : alpha
1 4 3 : species, atom, l
0.3678230245E-02 : alpha
2 1 2 : species, atom, l
0.2300745858 : alpha
2 2 2 : species, atom, l
0.2298197193 : alpha
Therefore, for La (atom 1) it is mostly FLL, whilst for Cu it is about 23% FLL? I don't have band structure right now but I can show you some results on the optical properties. They're attached
Do note that these are using different k-point grid and smearing (888, 0.01 vs 666, 0.008). I don't see a significant difference in the shape and position of the peaks, but there seems to be a bit more in Drude/intraband (I think) in lower energies. a-, b-, and c-axis components are solid, dashed, and dotted lines.
having 0.23 I assume its more AMF. However, do you have a ordered spin moment? Here, we have to be careful now about the magnetic state. Dependening if you have a PM solution or real AFM state the alpha value might be quite different. Hence, you should deside which system you want to describe
more the AFM insulator
or the PM metal
just you should check. From my experience in cuprates sometimes is difficult to get both stable solutions.
best
Michael
PS: By the way what is also nice to do is to set alpha by yourself and scan from AMF to FLL. You can do this by setting "readadu" and then setting any inbetween value. This is useful if you see a drastic turnover between the two approximation and to search for its microscopic origin.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Noted. I wanted to describe the AFM insulator, and so this should be a good alpha value. I admit it did take quite a while for it to converge.
Will try using <readadu>. Alpha values with different k-point sampling and swidth, based on my test, are quite similar, around 0.2 - 0.23. This should speed up the run right? Since Elk doesn't need to find the alpha.
Best,
Cenna
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Just you should get sure you get an ordered magnetic moment by checking INFO.OUT. Regarding the readadu, you do not obtain speed by reading in the alpha value since its calculations takes only seconds. Instead, reading in alpha gives you more control since you always have the same value. This might be helpful for example if you are interested how properties alter with the structure and thus freeze in alpha. However, in setting alpha by yourself, you also choose in principle a specific kind of orbital resolved exchange-correlation functional, which would be self-determined if alpha is computed in each cycle. To its end its a question which flavor you like in this respect.
best regards
Michael
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear Kay,
I understood that we can break spin symmetry for magnetic calculation with a small bfcmt or bfieldc.
Which routine should I take a look to understand the symmetry breaking by these fields ? I'm trying to understand the optimal ratio between bfcmt and bfieldc to study the extermal magnetic field effect.
Best,
bbbusybee
Hi!
Take a look at the routine 'seceqnsv'. The second-variational Hamiltonian is of the form
H = H_0 + sigma.B + H_so
where H_0 is the first-variational Hamiltonian (diagonal in the first-variational states) and H_so is spin-orbit coupling. The sigma.B term is the one your interested in: both bfcmt and bfieldc are applied like this in exactly the same way. The only difference is that bfcmt is in the muffin-tin, and bfieldc is applied everywhere.
The ratio shouldn't matter. Most of the time, the B-fields are used to break spin symmetry in a particular way.
Cheers,
Kay.
Hi Kay,
I'm doing a DFT+U spin-polarized calculation with the parameters below (incomplete, I'm actually doing La2CuO4)
mixtype 3 broydpm 0.6 0.01 stype 0 bfieldc 0.005 0.005 0.0 'Cu.in' : spfname 2 : natoms; atposl, bfcmt below 0.00000000 0.00000000 0.00000000 -0.10000000 -0.10000000 0.00000000 0.50000000 0.50000000 0.00000000 0.10000000 0.10000000 0.00000000Using an AMF double-counting with Slater parameters, this converges, but very slowly (about 1,5 days with 40 cores).
I wanted to try using U + automatically determined Yukawa screening length and an interpolation between AMF and FLL (i.e. dftu 3, inpdftu 5), however the calculation takes even longer (more than 2 days with the same setup).
I was thinking of ways to speed this up:
I'm a bit confused since the parameters are very dependent of one another.
Best,
Cenna
Dear Cenna,
your input is a little bit counteracting. You are applying a local magnetic field on Cu to enforce a AFM state and at the same time you also apply a global magnetic field, which favors a FM state. From my experience I would suggest you to do the followin:
1) remove bfieldc
2) increase the field at the atoms to a value between 1 - 5
3) set reducebf = 0.8 -- 0.9
4) set mixing to linear
this should speed up things a little bit. Actually, broyden mixing (for me in transitionmetal oxides) is not always really faster.
Regarding your third question
--> "Will there be any significant difference, especially in band structure and RPA results, between the AMF and the AMF-FLL Interpolation?"
This depends in my experience Cuprates tend to be more FLL. I would suggest here the following strategy first compute FLL and AMF and than you can run the mixing scheme (or all three at the same time if your cluster allows). There is a file written out (alpha.OUT) or something where elk tells you, for the FLL-AMF mixed scheme, what the actual interpolation parameter is. If it is away from 0 (meaning actually AMF) than you may can expect differences.
hope this clarifies
best regards
Michael
PS: It would be nice to post your observation here, what the actual alpha in the interpolation scheme is and if it makes a differnence.
Hi Michael,
Thank you so much for you input. Currently testing your suggestions and I'll share with you when it's done.
Is there any particular reason for using linear mixing instead of the default adaptive mixing?
I've done AMF and the FLL-AMF interpolation so far, here's the alpha values
Therefore, for La (atom 1) it is mostly FLL, whilst for Cu it is about 23% FLL? I don't have band structure right now but I can show you some results on the optical properties. They're attached
Do note that these are using different k-point grid and smearing (888, 0.01 vs 666, 0.008). I don't see a significant difference in the shape and position of the peaks, but there seems to be a bit more in Drude/intraband (I think) in lower energies. a-, b-, and c-axis components are solid, dashed, and dotted lines.
Best,
Cenna
okay,
alpha=0 means AMF
and
alpha=1 means FLL
having 0.23 I assume its more AMF. However, do you have a ordered spin moment? Here, we have to be careful now about the magnetic state. Dependening if you have a PM solution or real AFM state the alpha value might be quite different. Hence, you should deside which system you want to describe
more the AFM insulator
or the PM metal
just you should check. From my experience in cuprates sometimes is difficult to get both stable solutions.
best
Michael
PS: By the way what is also nice to do is to set alpha by yourself and scan from AMF to FLL. You can do this by setting "readadu" and then setting any inbetween value. This is useful if you see a drastic turnover between the two approximation and to search for its microscopic origin.
Hi Michael,
Noted. I wanted to describe the AFM insulator, and so this should be a good alpha value. I admit it did take quite a while for it to converge.
Will try using <readadu>. Alpha values with different k-point sampling and swidth, based on my test, are quite similar, around 0.2 - 0.23. This should speed up the run right? Since Elk doesn't need to find the alpha.
Best,
Cenna
Hi Cenna,
Just you should get sure you get an ordered magnetic moment by checking INFO.OUT. Regarding the readadu, you do not obtain speed by reading in the alpha value since its calculations takes only seconds. Instead, reading in alpha gives you more control since you always have the same value. This might be helpful for example if you are interested how properties alter with the structure and thus freeze in alpha. However, in setting alpha by yourself, you also choose in principle a specific kind of orbital resolved exchange-correlation functional, which would be self-determined if alpha is computed in each cycle. To its end its a question which flavor you like in this respect.
best regards
Michael