Convergence problem of SS-NEB

[Donna]

Dear Prof. Henkelman and users,

I am simulating the phase transition of MoTe2 from 2H to 1T'. Since the two phases have different unit cells, the SS-NEB is used. But it is difficult
to get atomic forces below a threshold such as 0.05 eV/Ang.

At first, I tried to do a rough calculation with IMAGES=5, EDIFFG=-0.1 and LCINEB=.TRUE.. But atomic forces are still large after 100 steps. Then
I tried to close the CINEB with IMAGSE=5, EDIFFG=-0.1 and LCINEB=.FALSE.. It worked and converged at a proper state. Using the CONTCARs of
the last step as the new POSCARs, I repeated the work with EDIFFG=-0.07 and LCINEB=.FALSE.. Everything was fine. But when I improved the convergence criteria with EDIFFG=-0.05 and LCINEB=.FALSE., it didn't converge again.

Here I attached the files that I used with EDIFFG=-0.05 and LCINEB=.FALSE.. Could you please give me some suggestion or possible solution to my problem? Thanks in advance.

Best regards,

Baojuan Dong
Institude of Metal Research, Chinese Academy of Sciences.

[graeme]

This calculation may be alright. There are, however, some unusual aspects to the geometry and reaction pathway. You have a 2D layered material where the distance of the cell in z is somewhat arbitrary. Additionally, the angles between the z-x and z-y directions are presumably soft degrees of freedom, since the layered material can rotate the atomic structure normal to the plane to accommodate changes in the cell degrees of freedom. What I am trying to say is that since you have relatively unconstrained degrees of freedom, the reaction pathway will also be unconstrained in terms of cell vs atomic degrees of freedom. This should not, however, effect a converged saddle point energy. Using a small time step should result in such convergence even if the pathway looks a little rough in terms of cell vs atomic motion.

You could also consider reducing the number of free variables in your calculation. If you can represent your initial and final states using the same cell (even if one is tilted), then you can use a regular NEB calculation in which the cell degrees of freedom are fixed. This should avoid the redundant coordinate problem and lead to smoother relaxation to the minimum energy path.

[Donna]

Dear Prof. Henkelman,

Thank you very much for your kindly suggestion! I have already tried a smaller time step with EDIFFG=-0.05, LCINEB=.FALSE. and TIMESTEP=0.01. But it didn't change anything. After 100 steps, the force was still large.

As you mentioned, the changed angles between the z-x and z-y directions may be the key problem in my calculation. Then I tried to change the constr_cell_relax.F and rebuilt VASP. In this way, I wished to not change the Z direction. But when I did NEB calculation, It didn't work. The Z direction was still changed. Here I attached the changed constr_cell_relax.F file.

Is there any way for me to unchange the Z direction during the optimizing process with ISIF=3? Could you please give me some suggestion on how to keep the Z direction fixed?

Thank you very much for your precious time!
Best regards,

Baojuan Dong,
Institude of Metal Researh, Chinese Academy of Science

[graeme]

Try a fixed cell calculation where just the atomic layers move.

[Donna]

Dear Prof. Henkelman,

Thank you so much for your suggestion!

I have already got a converged result. But I still wonder whether my calculation makes sense or not. Since the initial state and final state of my NEB reaction have different unit cells, all the images have different cells, too. But when I did the NEB calculation, I turned off the LNEBCELL with LNEBCELL=.FALSE. and ISIF=2. In this way, all the cells of the images could not changed during the process of calculation. I got the converged results. But I am not sure if the results are right or wrong. Is this the way that you want to tell me about "Try a fixed cell calculation where just the atomic layers move"? Looking forward to your reply!

Thank you very much for your precious time!
Best regards,

Baojuan Dong,
Institude of Metal Researh, Chinese Academy of Science

[graeme]

I think that you might try two follow-up calculations. First, see if you can get the initial and final states in the same cell geometry without changing the relative energy of the two. Then you can compare what you have to a fixed-cell calculation.

Second, try taking your relaxed NEB in which the cell changes (linearly, I assume) and turn on the variable cell (GSSNEB) flag. With a reasonable starting geometry you may find that convergence is smoother. I still feel that the system is formally under-constrained in the sense that changes in the x-z and y-z angles can also be represented by atomic motion. Unfortunately, we don't have a good way to pick-and-choose which cell degrees of freedom are included in the calculation, so it will be interesting to see if you can converge the GSSNEB starting from the path that you have.