Resturt from previous run in atk 2010

[alex_z]

Dear all,

I'm a new user. I started from 2010 version and know nothing about previous 2008 one. I'd like to ask you how to start a new job from a previous run or possibly from interrupted run in ATK 2010. What I'm trying to do is get I/V curve for 2 probe system with a small organic molecule between the leads. The leads and the molecule are composed of Carbon and Hydrogen with the total number of atoms about 100.

I used the bias loop script provided here: http://www.quantumwise.com/documents/tutorials/GrapheneJunctionDevice/XHTML/chap.current.html

Even though the script mentions the "use the self consistent state of  the old calculation as starting input", I have a strong suspicion that it only picks up the geometry configuration and previous calculator parameters with modified bias, but not the self consistent field state from the previous run. I tried to nlprint the "device_configuration" object in the manual mode and got the geometry of the system. The restarting from previous run is important because my calculations of current do not converge even when I raise the temperature up to 1000 (I used extended Huckel method), so the restarting from previous run could possibly solve my problem of convergence.

I found some tips how to restart from previous run in 2008 version (http://www.quantumwise.com/documents/manuals/ATK-2008.10/chap.tipstricks.html), but I guess it won't work in 2010. I searched the atk 2010 manual and didn't find anything like "runtimeParameters" or "checkpoint_filename" parameters.

In sum,

1. I would appreciate if anybody tells me a legitimate way to start my calculations from the previous or interrupted run in atk 2010 (or proves me wrong about my suspicion regarding the "device_configuration" functionality).
2. Could you suggest something how to improve the convergence apart form raising the temperature?

Thanks.

[Anders Blom]

The crucial line to making sure the new calculation starts where the old one finished is the keyword

Code

initial_state=device_configuration

Never mind what nlprint says :-) It's defined in a certain way when applied to each type of object, and so nlprint(configuration) prints the geometry. You actually don't have the option to restart from an interrupted run in ATK 10.8 because the self-consistent state is only saved to the NC file when the calculation finishes. We'll introduce checkpoint files soon.

[jdgayles16]

I would be very interested in convergence tricks without changing the temperature. I have a bias point that i have been trying to converge for days, and the energy has met the criterion but the hartree potential is giving me trouble.

Jacob 

[jdgayles16]

Code

| 122 E = -765.983 dE =  3.034982e-04 dH =  4.211894e-05                       |
| 123 E = -765.983 dE =  2.804950e-08 dH =  4.217083e-05                       |
| 124 E = -765.983 dE =  1.747253e-08 dH =  4.211418e-05                       |
| 125 E = -765.983 dE =  3.189939e-09 dH =  4.215133e-05                       |
| 126 E = -765.983 dE =  1.969966e-09 dH =  4.215458e-05                       |
| 127 E = -765.983 dE =  1.499757e-09 dH =  4.215810e-05                       |
| 128 E = -765.983 dE =  2.668173e-08 dH =  4.215502e-05                       |

these are my last few electronic iterations any body have any suggestions how to get this to converge its keep doing this and then jumps to 10^-3

[Anders Blom]

It's not the Hartree potential that doesn't converge it's the Hamiltonian. Let me guess, you have a DoubleZetaPolarized basis set? It might be overdetermined, and is shuffling some small amount of weight between orbitals, without affecting the energy. How does dH look before this? Maybe you can just use a slightly less stringent convergence criterion; as you see the energy is converged anyway, like 5e-5. Actually the default criterion in 10.8 is quite strict, probably much stricter than in 2008.10. So, I would suggest lowering the tolerance, and monitor the total energy so that you are comfortable with it at the end of the convergence.

[alex_z]

Thanks a lot Anders.

[jdgayles16]

Thanks, do you think changing the hamiltionian mixing will effect this

[Anders Blom]

There is a reason Hamiltonian mixing is default. Most other DFT codes would use density matrix mixing, but that's for periodic or finite structures. For open systems, it was found a long time ago that Hamiltonian mixing, contrary to popular belief at the time, works better.

So, yes, I do think that mixing on the Hamiltonian will have a beneficial influence on the convergence since all that's left is to get dH down, the system is physically converged.

And, in case you are already doing that, perhaps increasing the mixing parameter a bit might be a good idea. And, as mentioned above, since the total energy is converged, all is good, so perhaps you don't need such a strict convergence criterion as the default.

[Anders Blom]

I could reproduce the case of dH not converging but everything else being small in a much simpler system. In that case, it was related to the electrode being too short (3 layers was too little, making it 4 made everything converge nicely).