Dear all:
I am calculate some reaction with NEB method.
Firstly, I set EDIFFG to positive value, and obtain a MEP.
However, when change EDIFFG to negative value, I
find a difference energy barrier with a smaller value.
In my opinion, since the imagines are the unstable structures,
the force convergence is impossible in NEB calculations.
Is force convergence imperative in NEB caculations?
Thanks
Is force convergence imperative in NEB caculations?
Moderator: moderators
You should only use the force criteria (ediffg<0), for transition state and minimization calculations.
It is my view that the energy-based criteria (ediffg>0) should never be used. The problem with the energy-based criteria is that convergence is based both on the potential and the choice of optimizer. A slow optimizer will consider an optimizer converged before an aggressive one. Also, the change in energy at an optimization step depends upon the history of the optimization (for example the velocity in quick-min is cumulative, bfgs builds an inverse Hessian matrix, and conjugate gradients moves conjugate to previously optimized directions), so two optimizations of a structure to the same energy based criteria using the same optimizer could results in very different levels of convergence. Furthermore, I can see no reason not to use a force-based criteria.
The NEB algorithm optimizes the components of the force perpendicular to the band, and the spring forces along the band. These projected forces rigorously go to zero at convergence. The forces due to the potential along the band (along the unstable mode near the saddle) do not factor into the calculation.
It is my view that the energy-based criteria (ediffg>0) should never be used. The problem with the energy-based criteria is that convergence is based both on the potential and the choice of optimizer. A slow optimizer will consider an optimizer converged before an aggressive one. Also, the change in energy at an optimization step depends upon the history of the optimization (for example the velocity in quick-min is cumulative, bfgs builds an inverse Hessian matrix, and conjugate gradients moves conjugate to previously optimized directions), so two optimizations of a structure to the same energy based criteria using the same optimizer could results in very different levels of convergence. Furthermore, I can see no reason not to use a force-based criteria.
The NEB algorithm optimizes the components of the force perpendicular to the band, and the spring forces along the band. These projected forces rigorously go to zero at convergence. The forces due to the potential along the band (along the unstable mode near the saddle) do not factor into the calculation.
It is true that the real force on each image does not go to zero - is the the projected force plus the spring forces that goes to zero. But this (projected + spring force) is exactly that force that is used to determine convergence with ediffg.
In other words, everything should be working just fine. Set ediffg to something like -0.01 and the NEB will stop when the NEB forces drop below this value.
In other words, everything should be working just fine. Set ediffg to something like -0.01 and the NEB will stop when the NEB forces drop below this value.