The problem about SCF convergence.

[kevin]

Hi, all

Are there some artifice to enhance the SCF convergence of ATK?
Or who can provide me some documents about the SCF convergence of ATK?
It is not converged even after 300 SCF steps.
My system is consiste of
Ni electrode{with 3*3 and 4 layer(left) + 5 layer(right) surface} + Cu wire(5 atoms).

Thanks in advance,
Kevin

[Nordland]

Hey Kevin.

Sadly NEGF is not so stable as usual DFT, and hence it can often require a little know-how to get the systems converging
in the first go. However there is a pretty good guide to get you started under the tutorial section on the quantumwise.com page.
You can go directly to the page by pressing this link http://quantumwise.com/publications/tutorials.

It is called Convergence Tricks and it has helped me alot - however if you still have problems converging your system, you could post it here and I could try to see, if I know any magic buttons to turn.

[kevin]

Hi Nordland,
Thanks for your reply.
These days i have tried new attempt follow your suggestion about SCF convergen for two-probe.
But i have not get way to solve such problem.

Could you give me some advice?
I attached the input file and output file.
Although the calculation is not complet from the output file,
i can not find any convergency tendency.

Thanks in advance,
Kevin

[Nordland]

Hey Kevin!

It seems you have pushed the right buttons - however I will try to take a look at as soon as I get a free slot in my calculation machine. What is the most striking thing,
that it is the equivalent bulk part of the calculation that has trouble converging, and that I have only seen when the geometry is wrong. However when I inspect
your structure in VNL I can't seem to find any defect or any kind.

The only thing I am bit uncertain about, is the basis set, which might be too small for dealing with vacuum around the junction, perhaps I would try going for 10,10,10 kpoints and use a better basis instead for your choice.

Best regards,
   Nordland.

[Anders Blom]

Generally speaking, I would not expect reliable results for Ni unless you run a spin-polarized calculation.

About k-point, 10 in the z-direction seems a bit on the small side, I would use (10,10,25) instead, perhaps.

The system looks very interesting, I hope you get to terms with it quickly so you can get some interesting results!

[Nordland]

Hey Kevin.

I got a little hooked on your problem and tried to understand it. I tried to due some simulations on the systems that look like your system.

First of all I tried atomic nickel and atomic copper with single zeta basis and they both works fine.
Then I tried making a nickel bulk system that looks like the one you are modelling - just without the chain, and it works really well with SingleZeta.
The same for a copper chain worked perfectly, so I tried to combine nickel and copper into a sample system. Making them into a diatomic wire,
in order to understand what makes copper and nickel so unhappy when put together and I found that the way to make them work together is
to use a slightly higher mesh-cutoff.

An my conclusion is then:
Nickel and Copper has a very small difference in the atomic configuration and hence the basis set is somewhat similar, this lead to the fact that when determining which orbitals should be occupied or not, can be tough for the mixing algorithm to settle. Therefore if you increase your mesh-cutoff to 200 or 250 Ry, you can properly be more loose on the other parametes and still get convergence.

I have attached my three test scripts that I used for tracking this one down.

test_1.py is the Nickel bulk system, that converges straight away.
test_2.py is the Nickel/Copper chain with default mesh-cutoff, that does not converge in 100 steps.
test_3.py is the same as above, but with a mesh-cutoff of 200 Ry, which converges in 14 steps.

I hope this help your further.

Edit: Forgot to attach the scripts.