too hard to converge when calculating the transmission spectrum

[lknife]

Dear all,

I am calculating the transmission spectrum of a device bias by bias to get the IV curve. The method I am using is Extended Huckel method and I am using ATK 2015.1. I met several problems about the convergence of the calculation.

Firstly, I have to change some parameters when the bias voltage changes, such as k-point sampling, damping factor and history step, to make the calculation converge. I am not sure if these adjustments will have some influence on the current value calculated.

Secondly, I have big problem about the convergence of one of the calculation. The transmission spectrum calculation converged easily at bias -0.21V, but it cannot converge at bias -0.24V. I have done almost all I can do trying to make it converge.
These attempts include:
1. increasing the k-point sampling from (1,9,100) to (1, 15, 150), even to (1, 20, 200);
2. adjusting the damping factor and history steps with (0.05, 25), (1, 20)(default), (0.3, 12);
3. using the converged configuration under bias -0.21 as the initial state;
4. increasing the maximum iteration steps up to 1500

All of the above attempts were failed. Attached is the .py file for one of the failed calculation. Since the .log file is very large, I cannot post it here.

Thirdly, I heard from the turotial/Forum that I can increase the electron temperature which may be help to the convergece. But I am still not sure if it is all right if the IV curve contains current calculated at different electron temperture.

Can anybody here help me with these problems? Thank you very much for your kind help!

[Ulrik G. Vej-Hansen]

Have you tried combining them? You should always use the converged state at the previous bias as the initial state.

The number of k-points in the c-direction is quite important, and actually not that computationally demanding, so I would suggest using 200 there in all cases.

1000 iteration steps should be more than enough, so I would focus on other options before increasing them. However, one strategy here is to have a rather large number of history steps and a low damping factor. This can improve convergence, but it will also require many steps.

I am not sure what you have available in 2015.1, but changing the Poisson solver can sometimes help convergence.

The electron temperature should not affect the results much, but it certainly can, so you need to check that.

Finally, you should also verify that the lower bound of the contour integral is low enough, and you could also try increasing the number of points here.

[lknife]

Thank you very much for your kind reply and valuable advices. Now the calculation has been converged.

Thank a lot again for you help!

[Ulrik G. Vej-Hansen]

Very good, I am happy to hear that it worked out