QuantumATK Forum
QuantumATK => General Questions and Answers => Topic started by: Hasan Sahin on December 28, 2008, 12:07
-
After Electrodes Calculation and Equivalent Bulk Calculation (Initial Density for TwoProbe), during the TwoProbe Calculation
ATK sometimes gives the error
# Pulay mixing inversion failed. Using only last step.
and can not converge. How can I solve this ?
-
This error can happen in two cases:
1) Either the results is soo well-converged - like 1e-16 - that the linear algebra behind the pulay mixing becomes numerical unstable
2) Or you results are soo messed up, that everything is lost :)
Since you are saying that, it is doing the TwoProbe SCF I would assume, and not during the equivalent bulk part, that
it is due to issues in the interface between the electrode and the scattering region. But of course I can't be sure, and
it is only my guess.
However when this error comes and you are not the the first guess, I would suggest looking at geometry and everything is correct,
the I would increase the accuracy of the energy contour integral parameters f.x by lowering the lower bound and the number of circle points.
-
I increased history steps to 12, lowerbound to 4 and circle points to 50. ATK does not give the error but convergence still takes too long (and it may give error). I attached the geometry of the system. Does geometry script have mistake ?
-
The geometry looks cool, and there is no critical errors at all.
I have tried to start a sample calculation on it on my laptop - trying out some parameters to see if I get the same problem as you.
-
Now I am thinking about electrode-device atom matching criterion... I hope one of us manage to calculate an I-V...Thank you for time and effort.
-
I have now converged the system with Single Zeta and SingleZetaPolarized. I had some trouble using DZP, but I will try again now,
once this calculation is done, then I will post my parameters :)
-
At the same time I did it. It is impossible to converge it with DZ basis. It seems that voltage steps between two scf calculation must be smaller than 0.01 volt. I tried 0.02 volt but it can not find the initial density of next step by using scf of 0.00 volt.
This is a trial calculation with a very simple system, it seems that it is impossible to get I-V characteristics of larger systems. I am discouraged.
-
I know the feeling! I have learn the hard way how much tougher it is to get NEGF to converge compared to usual DFT :|
However there is hope. I have done IV calculations on huge graphene sheets with 6000 atoms - so it is possible :)
And it was even with a gate voltage and I tried to model that these graphene sheets could be used a field-effect transistor,
and with good results.
The best way to get started on cracking these problems is for me, that starting out by a trial system, that involves the same chemical elements,
and geometries types as the target system. Then it is fast to try a few things and learn what makes this calculation converge fast and rock-solid.
Almost always can this insight be transfered to the target system right away.
I know there exist a little advanced tutorial on getting twoprobes to converge better and safer, but I dont have it.
Perhaps you can get it from QuantumWise if you write to them.....
-
I got it working for DZP as well :)
-
My calculations with SZ and SZP are ok.
1-) I am curious about your DZP calculation (I couldnt get any I-V value with DZ and DZP).
1a-) How many different I-V values did you calculate?
1b-) Which parameters did you changed in your DZP calculation script?
2-) Do you have a published paper or a link about your graphene FET ?
-
Hello Hassan,
the geometry you posted attracted me also since i'm also doing some calculations with carbon structures. i have some graphene geometries also on my web page: geocities.com/carbon89 if you r intereseted. but like u I did not achieve a I-V characteristics with polarized basis sets before. by the way, on which structures/subject are you working on??
-
1-) I am curious about your DZP calculation (I couldnt get any I-V value with DZ and DZP).
1a-) How many different I-V values did you calculate?
1b-) Which parameters did you changed in your DZP calculation script?
I tried a couple of things to be sure it converged, it has run smoothly, but I killed the calculation this morning when it had
done up to 0.45 volt in steps of 0.05 volt (10 seperate calculations), since I only have my laptop for doing these calculations :)
- 25 history steps with 0.05 diagonal mixing parameter
- 200 Ry mesh-cutoff ( most likely not needed )
- 1000 max iterations ( not needed since all voltages converged in 20-40 steps )
- (1,1,100) kpoints
- 600 Kelvin electron temperature(Improves alot, but does not do it alone)
- 0.05 eV real axis infitesimal(Improves alot, but does not do it alone)
- DensityMatrix electrode constraint(Best thing in ATK 2008.10 - apart from the speed-ups)
- 0.05 Volt steps
With these parameters It runs smoothly for DZP, and I will post the IV-curve once I get the data processed.
2-) Do you have a published paper or a link about your graphene FET ?
I have not published the results :| - but let me see if I can find the scripts again, and then I can share them with you.
-
Hello Hassan,
the geometry you posted attracted me also since i'm also doing some calculations with carbon structures. i have some graphene geometries also on my web page: geocities.com/carbon89 if you r intereseted. but like u I did not achieve a I-V characteristics with polarized basis sets before. by the way, on which structures/subject are you working on??
Hello carbn9,
Especially I am working on doped graphene flakes and graphene ribbons.
-
1-) I am curious about your DZP calculation (I couldnt get any I-V value with DZ and DZP).
1a-) How many different I-V values did you calculate?
1b-) Which parameters did you changed in your DZP calculation script?
I tried a couple of things to be sure it converged, it has run smoothly, but I killed the calculation this morning when it had
done up to 0.45 volt in steps of 0.05 volt (10 seperate calculations), since I only have my laptop for doing these calculations :)
- 25 history steps with 0.05 diagonal mixing parameter
- 200 Ry mesh-cutoff ( most likely not needed )
- 1000 max iterations ( not needed since all voltages converged in 20-40 steps )
- (1,1,100) kpoints
- 600 Kelvin electron temperature(Improves alot, but does not do it alone)
- 0.05 eV real axis infitesimal(Improves alot, but does not do it alone)
- DensityMatrix electrode constraint(Best thing in ATK 2008.10 - apart from the speed-ups)
- 0.05 Volt steps
With these parameters It runs smoothly for DZP, and I will post the IV-curve once I get the data processed.
2-) Do you have a published paper or a link about your graphene FET ?
I have not published the results :| - but let me see if I can find the scripts again, and then I can share them with you.
Ok I will try with these parameters and upload the IV curve here . A calculation including 6000 atoms sounds good :))). Thank you for your time and fast informations.
-
Finally I got around to getting some computer power to do the current calculation, as I had only did the SCF part first :)
For comparison I also calculated the linear response current to see the effect of the self-consistent response.
I have attached both the script and example of the IV curve for this system.
Cheers!
-
Hi Nordland,
Could you please briefly explain the difference between SCF current and linear response current?
Serhan
-
Well I can give it a try.
Linear Response Current
When you calculate the linear response current, you only perform a self-consistent calculation of the system at zero bias,
and from this calculation you calculate the transmission spectrum T(E). The current is then calculated by performing
an integral between the two chemical levels (adjusted for the voltage) of the electrodes of the transmission and
this you is linear response current.
- Alot faster! (only one SCF calculation )
- Good convergence ( no trouble converging calculations with a bias )
- Limited domain of good accuracy
- Unable to show advanced features of the current curve - like spin-torque in FeMgOFe
SelfConsistent Current.
The effective potential of the twoprobe takes into account that the bias applied has given a voltage ramp in the potential, which
leads to a new hamiltonian, which leads to a new NEGF function, which leads to a new density matrix, which produces a new effective potential.
So this process is continued until there is convergence - just like normal DFT. This selfconsistent process is the essences of ATK, in my eyes.
Once converged, the transmission spectrum is calculated for this state, and the same integral as above is performed and the current is
calculated.
- Very high accuracy
- Results takes into account the perturbation of the system created by the bias.
- Can reveal key-features of the systems
- Calculations time is substantial longer ( many SCF loops)
- Prone to convergence troubles
Which to choose?
It is hard to put a 100% guideline for which to choose, but if I have to an investigation of some materials,
I always first use the selfconsistent approach to see for some relevant sample systems, and compare the IV-curve with
the one obtained from the linear response. If the agreement is acceptable within the bias windows of my study,
then I stick with the linear response current.
However if the system is Magneto-tunnel Junction and I want to model the spin-current through the insulator layers, then I always use the
selfconsistent current.
I hope this answered your question :)
-
Thanks Nordland. It was very very helpful ;D