About the convergence of the transmission spectrum calculation

[lknife]

Dear experts,

When calculating the IV curve of a device (I am using the ATK2015 and ATK2016, Huckel method), if I used the "IVCurve" block directly, sometimes it was found that the transmission spectrum did not converge at some specific bias values. Thus, I had to modify manually the parameters of the calculation, such as the damping factor, the history step and the Poison solver, and redo the transmission spectrum calculation bias by bias, which is very inconvenient. In some cases, it was very hard to converge (>1000 iteration steps). However, when changing to a suitable parameter set, it converged within 50-100 steps.

The are two questions about this problem:
(1) Is there any problem if the points of the IV curve were obtained using different parameters?
(2) Is there a better IVCurve block that can adjust the parameters automatically when it's hard to converge?

Thanks a lot for your kind help!

[Petr Khomyakov]

(1) Is there any problem if the points of the IV curve were obtained using different parameters?

It really depends on which parameters are different. Varying damping factor, the history step and the Poison solver is meant to find a faster way to converge to the same solution of your problem; this solution should not  depend on the choice of these computational settings, at least within some numerical noise.

If you use, for example, individual k-point sampling or density mesh cut-off for different  bias voltages applied, then the current is not computed with the same level of accuracy for these different bias voltages. In this case, you have to do the entire IV curve calculation for the most accurate  parameter set, i.e., with the most dense k-mesh and the largest density mesh cut-off value.

(2) Is there a better IVCurve block that can adjust the parameters automatically when it's hard to converge?

Computational settings are to be adjusted manually. In principle, you can write a script to do calculations in a loop using different computational parameter sets.

[lknife]

Dear Petr Khomyakov,

Thank you very much for your kind help! It's really helpful!

lknife

[lknife]

Another question:

I am studying the gate effect of a device. I have previously got some IV plots at different gate voltages with relatively less k-point sampling (less accuracy). However, it's hard to converge at some gate voltages. So I had to modify the parameters to improve the convergence, including the number of k-point sampling.

I don't want to redo the IV calculations with less accuracy since it's also very time consuming. But I am not sure if it is possible or reasonable to compare two IV curves obtained with different accuracy. Could you please help me with this question?

Thank you very much again for your kind help!

[Petr Khomyakov]

In principle, you may still use different k-point meshes at different bias voltages, but then you have to make sure that your results are converged with respect to the k-point density (sampling).  Converged means that the current at any given voltage changes within some tolerance value when you increase the k-point density. And you decide on the tolerance value yourself, based on what is the target accuracy for the quantity of interest, e.g., the current in your case.

As a matter of fact, the numerical accuracy of your calculation is estimated from this kind of convergence tests with respect to various computational settings such as k-point sampling, for instance. 

[lknife]

Thank you for your reply, Mr. Petr Khomyakov!

I can make sure that all the calculations are converged with the same tolerance. If so, can I compare two IV plots obtained with different parameter sets directly?

[Petr Khomyakov]

If they are converged, I would say that it is meaningful to compare them.

[lknife]

Thank you very much for your help!