Hi Prof. Graeme,
1. To get the relaxed structure for a dynamical matrix calculation, should I really have to set EDIFF to such a small value as 1.0E-7?
2. When I use the relaxed structure from question 1 to perform a dynamical matrix calculation in order to get the vibrational frequecies, should I really have to set EDIFF to such a small value as 1.0E-7?
I ask two questions above because I have tried several jobs and found that it was so time consuming to get the calculation converged if I set EDIFF to 1.0E-7, especially for large systems. Is it acceptable that I set it to 1.0E-6 or even 1.0E-5? Will there be a big error bar for calculated vibrational frequencies between 1.0E-7 and 1.0E-6 or even 1.0E-5?
Thanks a lot.
Questions about EDIFF for DynMat calculation
Moderator: moderators
Re: Questions about EDIFF for DynMat calculation
Just to make sure that it's clear, EDIFF is the electronic convergence and controls the accuracy of the forces.
1) You can get a well converged structure using a more standard value for EDIFF of 1e-4 or 1e-5.
2) For a dynamical matrix calculation, you should have accurate forces because you are doing a finite difference approximation on the force. In theory you could push EDIFF larger than 1e-7 and make larger displacements, but in my experience, this doesn't make any sense. It is not so expensive to reduce EDIFF to 1e-7 or 1e-8 and get accurate forces so that you can make small finite difference displacements on the order of 1e-2 or 1e-3 Ang.
It is surprising that there is a significant cost reducing EDIFF from 1e-5 to 1e-7. Perhaps you are using systems that are much larger than I am used to. But anyway, the key is to make sure that your calculated modes are insensitive to a change in EDIFF and the size of the finite difference step. This is how to check for convergence -- if the modes change significantly, the results can not be trusted.
1) You can get a well converged structure using a more standard value for EDIFF of 1e-4 or 1e-5.
2) For a dynamical matrix calculation, you should have accurate forces because you are doing a finite difference approximation on the force. In theory you could push EDIFF larger than 1e-7 and make larger displacements, but in my experience, this doesn't make any sense. It is not so expensive to reduce EDIFF to 1e-7 or 1e-8 and get accurate forces so that you can make small finite difference displacements on the order of 1e-2 or 1e-3 Ang.
It is surprising that there is a significant cost reducing EDIFF from 1e-5 to 1e-7. Perhaps you are using systems that are much larger than I am used to. But anyway, the key is to make sure that your calculated modes are insensitive to a change in EDIFF and the size of the finite difference step. This is how to check for convergence -- if the modes change significantly, the results can not be trusted.
Re: Questions about EDIFF for DynMat calculation
Hi Prof. Graeme, thank you so much for your reply. Now I know that setting EDIFF to 1.0E-7 is unneccesary to get the optimized structure, but is neccesary for dynamical matrix calculations.
I am calculating a system with a molecule adsorbing onto a 48-atom Pd slab. I think it's a big system for dynamical matrix calculation, especially when I have to set ISYM=0.
I am calculating a system with a molecule adsorbing onto a 48-atom Pd slab. I think it's a big system for dynamical matrix calculation, especially when I have to set ISYM=0.
Re: Questions about EDIFF for DynMat calculation
That is not so large. You will probably find that you don't need to displace all of the metal atoms. You can displace the molecular degrees of freedom, and then add the nearby metal atoms to see if the modes change.