Problem with convergence

[weiyidan]

Dear professor,
Thanks for your help all the time.
However, i have to ask for help again. When i relax the structure, I found it hard to converge. I tried IBRION=1 or 2 which cannot satisfy the standards (EDIFFG=-0.02). And the energy vibrated significantly. In the past, you told me that the energy should always decline and choose a smaller timestep with IBRION=1. But it seems that these does not work. Can you tell me how to solve it?
Yidan Wei

[weiyidan]

Here is the attachments. Thanks for your help.

[graeme]

Try using IBRION=3 and look for a steady drop in the force.

[weiyidan]

Dear professor，
Thanks very much，i will try today. I am very appreciate your knowledge and your kindness, can i transfer to your university? I want to learn more and study further with you.
Yidan Wei

[weiyidan]

Dear professor,
This is the first time for me to use IBRION=3. Therefore, i do not know which method should be used. Is the damped molecular dynamics mode or the velocity quench algorithm?
And Can you simply interpret the theory and the reason of divergence with IBRION=1 or IBRION=2？
Thanks very much and best wishes.
Yidan Wei

[graeme]

IBRION=1 and 2 both rely on calculations of curvatures. If there is noise in the force it can make these methods erratic. IBRION=3 is a first=order method, so it tends to be more stable but ultimately slower. Using POTIM=0.1 (quickmin) is generally stable. If the forces are particularly high, you can drop POTIM to about 0.02.

[weiyidan]

Dear professor,
Thanks for your help. And can i be a member of your group as a exchange student for one or two years？
Yidan Wei

[weiyidan]

IBRION=3 and POTIM=0.1 does not seems to work. How can i solve it? This obstacle makes me crazy =.=

[graeme]

You need to increase the accuracy of your forces by lowering EDIFF. In the attached calculation, I needed to reduce the mixing parameters. With more accurate forces, any of the optimization algorithms should work.

initial.tar.gz

(4.22 MiB) Downloaded 10460 times

[weiyidan]

Wow, it is really amazing. I appreciate your help very, very much.
How can you get the idea？ Can you tell the way of solving this kind of problems？I only know three ways to solve the problem of divergence.
First of all, it is needed to lower the ediff than that of unreasonable starting guess.
Secondly, reduce the POTIM when the energy vibrate very much, and achieving the continuous descent of energy.
Thirdly, change a algorithm.
Concerning to the present problem， I found the energy vibrated. Therefore, I reduced the POTIM and change a algorithm with a lower ediff. However, it does not work.
At the same time, i found the steps of electronic relaxation was so many. Is this the point causing a divergence?
Thanks for your help very much.
Yidan Wei