Clarification about previous answers

[cwandtj]

In previous answer for other users' question, it is explained that:

"When I talk about forces being accurate enough near the minimum, I'm referring to the errors in the forces which are controlled by the electronic convergence criteria, ediff. When the forces are larger than the errors, curvatures can be determined with sufficient accuracy. Near the minimum, the errors are a higher portion of the force. At this point, you can either go back to quick-min (or fire), or decrease ediff. Your value of ediff=1e-8 is extremely tight. Using a value of 1e-5 will generally speed up your calculations without causing problems unless you are trying to get forces down to ~0.05 eV/Ang."

Here, what do you mean by "the errors are a higher portion of the force"? Also, why do we need to go back to quick-min? It is explained that quick-min is better for the case that "stable far from the minimum where forces are high".

Also, in normal static calculations, I use EDIFFG=-0.01 eV/Angs. However, based your explanations, EDIFFG=-0.1 eV/Angs is enough. So, "generally", can we consider that EDIFFG= -0.1 eV/Angs is really enough? This is because I experienced that convergence in NEB is relatively difficult than normal DFT calculations and maybe keeping EDIFFG value from static calculations is not a good idea and not that efficient in determining reaction barrir.

Thank you for your help in advance.

Chan-Woo

[graeme]

The second order optimizers (CG and LBFGS) are generally more efficient than a first order optimizer such as quick-min. There are two cases in which this is not true (1) when the force are very high and the system is far from a harmonic region and (2) very close to the minimum if the force are inaccurate.

There is always some numerical error in the vasp forces. These are controlled by ediff. If these errors are on the order of 0.01 eV/Ang, and you are converging to a tolerance of 0.01 eV/Ang, the error will be a very high fraction of the force. Then, you can not make an accurate harmonic approximation. You may have seen a bracketing error in CG, or oscillations when using LBFGS near a minimum. You can fix this by tightening ediff or switching to quick-min, since it is less sensitive to errors in the force.

Converging forces to 0.01 eVAng is a good idea. In my experience, 0.1 eV/Ang is not enough. For an NEB calculation, if you only care about the saddle energy, then you only require that the saddle have this force criteria.

[kai]

[quote="graeme"]The second order optimizers (CG and LBFGS) are generally more efficient than a first order optimizer such as quick-min. There are two cases in which this is not true (1) when the force are very high and the system is far from a harmonic region and (2) very close to the minimum if the force are inaccurate.

There is always some numerical error in the vasp forces. These are controlled by ediff. If these errors are on the order of 0.01 eV/Ang, and you are converging to a tolerance of 0.01 eV/Ang, the error will be a very high fraction of the force. Then, you can not make an accurate harmonic approximation. You may have seen a bracketing error in CG, or oscillations when using LBFGS near a minimum. You can fix this by tightening ediff or switching to quick-min, since it is less sensitive to errors in the force.

Converging forces to 0.01 eVAng is a good idea. In my experience, 0.1 eV/Ang is not enough. For an NEB calculation, if you only care about the saddle energy, then you only require that the saddle have this force criteria.[/quote]

occillation using LBFFS near a min...

that is probably what i am suffering.....right now i am using ediff=1e-4....would 1e-5 be tight enough?

btw, how to only define Saddle point with higher force criteria?

[graeme]

It could be the accuracy of the force. The easiest way to know is to try it.

You might could try lowering INVCURV if that doesn't work.

I don't quite understand the last question. You can set the accuracy of the force and the force criteria separately for a saddle point search as you would for a minimization.

[lgxyz]

"......Using a value of 1e-5 will generally speed up your calculations without causing problems unless you are trying to get forces down to ~0.05 eV/Ang."

I canot understand this sentence very well. Does it mean that if ediffg is set to -0.05, 1e-5 of ediff will cause problems?
And if ediffg=-0.05, should the value of ediff be more rigid?
Many thanks!

[kai]

[quote="graeme"]It could be the accuracy of the force. The easiest way to know is to try it.

You might could try lowering INVCURV if that doesn't work.

I don't quite understand the last question. You can set the accuracy of the force and the force criteria separately for a saddle point search as you would for a minimization.[/quote]

thanks, i tried and it didnt improve the situation. now i am turning to incurv. could you give me a little bit hint on its physical meaning?

[graeme]

The INVCURV value species a lower value of the inverse curvature for the system. This is needed to initiate the approximate inverse Hessian in the (l)bfgs method. Basically the first step is a steepest descent step: dR = F * INVCURV. If INVCURV is small, the step will be more conservative. You can estimate an appropriate value from the maximum curvature in your system. A value of 100 eV/Ang^2 is very high so that INVCURV=0.01 is low; 0.1 is typical.

[kai]

thx, Graeme