Author Topic: Convergence problem of Heusler alloy (Read 7900 times)

phylyh · « **on:** August 12, 2010, 14:18 »

Dear colleagues

Is there anyone has done the spin-polarized transport calculation of Heusler alloy successfully? I have tuned any parameter I know to converge this system, but always failed. The INPUT file and OUTPUT file of parallel calculation are included. Hope someone can help me.

Thanks for your help

Anders Blom · « **Reply #1 on:** August 12, 2010, 16:48 »

I think this system might converge if you give it more iterations, say 200.

However, you have another problem. Your parallel setup is not working properly, which you can see from the fact that each line in the log file appears 4 times (guessing you run mpiexec -n 4!).

The reason in obvious in fact from the first line: you are using OpenMPI, which is not supported by ATK. You must use MPICH2.

Well, you can run, as you are doing, it will not crash (yet), but you are not getting any benefit whatsoever from the parallelization (i.e. it's running on the same speed as in serial), and there's a risk of messing up the NC file when all the nodes think they are masters and try to write to the same file (at once).

phylyh · « **Reply #2 on:** August 25, 2010, 04:46 »

Thanks for the reply!

This system still can not converge after 200 cycles and output file is included.

Now, I just want to converge this system and is there anyone can give me any advice?

Anders Blom · « **Reply #3 on:** August 25, 2010, 07:51 »

Our general experience for two-probe systems is that Hamiltonian mixing (default) is more appropriate than DensityMatrix, as you are using. Stick more to default parameters and give it another try!

nori · « **Reply #4 on:** August 26, 2010, 02:44 »

Dear phylyh,

Could you try attached?
I could get your system converged after 23 loops with it

I'm not sure what is the main cause about this issue because I modified several things from yours but anyway it would help you.

baizq · « **Reply #5 on:** February 12, 2011, 06:16 »

Dear nori,

Thanks for your help with our problem. I have tried your script and convergenve is really achieved after 23 loops. I have to say that your modified script is a great support to our work because we struggled to tune the parameters but the system was always failing to converge.

I also have a question here: why do you set the initial elemental spins as those in your script? How did you determine the values, such as 1.0 for Cr, 0.1 for Co, -0.13 for Si and so on? Since after convergence, I have to calculate the transmission properties for both antiparallel and parallel cases, and I have no idea how to set the initial spins for each kind of element involved, could you please kindly give me some suggestion on this issue?

Thanks!
baizq

nori · « **Reply #6 on:** February 14, 2011, 05:35 »

Dear baizq,

Quote

Thanks for your help with our problem. I have tried your script and convergenve is really achieved after 23 loops. I have to say that your modified script is a great support to our work because we struggled to tune the parameters but the system was always failing to converge.

It's my pleasure

Quote

I also have a question here: why do you set the initial elemental spins as those in your script?

I determined each initial scaled spin from spin polarizations of bulk Co2CrSi and Cu2CrAl.

Quote

I have to calculate the transmission properties for both antiparallel and parallel cases, and I have no idea how to set the initial spins for each kind of element involved, could you please kindly give me some suggestion on this issue?

In anti-parallel case, indeed it's not obvious.
One candidate is that spin reversal occurs at the center of Cu2CrAl.
How about attached?

baizq · « **Reply #7 on:** February 20, 2011, 07:23 »

Dear nori,

Many many thanks for your help.

I have tried the script for antiparallel case. It converges after 56 loops. Really appreciate your help.

baizq

baizq · « **Reply #8 on:** March 7, 2011, 08:31 »

Dear nori,

I have another two questions...

How to do a bulk calculation to find the spin value as initial guess...

And will the initial spin converge to the correct spin configuation after self-consistent iterations? For example, the central spacer Cu2CrAl is a non-magnetic material. Can we get it as non-magnetic after setting a initial spin and conducting scf run?

Thanks very much. (*^___^*)
baizq

nori · « **Reply #9 on:** March 7, 2011, 10:40 »

Quote

How to do a bulk calculation to find the spin value as initial guess...

Examples are attached.
You can obtain spin polarization of each atom by executing "co2crsi_bulk.py" and "cu2cral_bulk.py" and calculating the difference between up spin charge and down spin charge.
And then, you can estimate the adequate initial guess, for instance, for Cr is (spin polarization/4).
http://www.quantumwise.com/documents/manuals/ATK-2008.10/ref.electrondensityparameters.html#ref.electrondensityparameters.description

Quote

the central spacer Cu2CrAl is a non-magnetic material

It may be true in real, but ATK 2008.10 predicts that Cu2CrAl has also spin-polarization.
I doubt that the electronic structure of Cr can't be treated properly in ATK 2008.10.
So, I strongly recommend you get upgrade of ATK to 10.8 or 11.2 because this issue is fixed in new ATK.
http://quantumwise.com/forum/index.php?topic=667.msg3531#msg3531

baizq · « **Reply #10 on:** March 8, 2011, 03:02 »

Hi nori,

Thanks very much for your scripts and suggestion.

I will try ATK 10.8.

baizq

QuantumATK Forum

News: