QuantumATK Forum
QuantumATK => General Questions and Answers => Topic started by: m_spin on October 13, 2023, 05:50
-
Hello,
I am using QuantumATK to model a 2D tellurium double-gate MOSFET, and I am running into some trouble with NEGF convergence when trying to obtain IV-characteristics. I have been working through the solutions discussed in the NEGF convergence guide article: https://docs.quantumatk.com/technicalnotes/negf_convergence_guide/negf_convergence_guide.html, however, my convergence issues still persist. I’ve attempted to modify the channel and electrode length, experiment with the self-consistent loop parameters, and increase k-point grid density. Unfortunately, these haven’t improved the convergence issues significantly. Typically I can achieve convergence at the zero bias condition, though it takes quite a while to converge, and I haven’t been able to achieve convergence for the finite bias condition.
I’ve attached my atkpython script and log files for one of my attempts. I’m using the LCAO Calculator and running an IV sweep using the ivcharacteristics study object.
I’m still quite new to the QuantumATK software, so it’s also quite possible I’ve made a beginner’s error somewhere during the simulation setup.
It may also be worth mentioning that the MOSFET structure I’ve attached, which does not converge, has a 9 nm gate length. Previously, I was able to perform a fully converged IV sweep on a MOSFET device with a shorter 5 nm gate length. I’m not sure why increasing the gate length has resulted in these NEGF convergence issues. The same convergence issues appear when I try to include a gate underlap in the MOSFET structure.
If any assistance could be provided to resolve this convergence issue, it would be greatly appreciated!
-
Replying to attach the log of the bias point that did not converge.
-
Parameters generally look ok, and it's not an outright failure, it's just that convergence seems slow. You mentioned playing with SCF parameters, but did that include history steps (reduce to 10) and mixing parameters (decrease, or actually increase for more aggressive convergence)?
Otherwise I am thinking perhaps the doping profile changes too abruptly, you might need to taper it with a smooth transition from doped to undoped.
-
Thanks for your response!
I have tried to increase history of steps to 25 and 30, but I haven't tried decreasing the value as you suggested. I also haven't tried a smooth transition for the doping profile, so I'll see if that improves convergence times.
Since as you mentioned, convergence seems slow rather than outright failure, I suppose increasing the number of SCF iterations beyond my current setting of 400 could also eventally take the system to convergence?
-
Sure, increasing the steps might work, but it's kind of pity it should take so long, so in case some trick like mixing helps it converge in 100 steps instead, you will save a lot of time