NEB -> CI-NEB -> Dimer (forces continually increasing)

[arosen]

I am trying to find a TS involving H-abstraction. I optimized the proposed initial and final state. I then did an NEB calculation. The MEP looked appropriate -- the endpoints are the local minima and there is a single local maximum between them. I tried to use the dimer method from the interpolated peak, but it wasn't converging to a desirable TS. Instead I ran a quick CI-NEB calculation from the original NEB run at a high EDIFFG of 0.1 eV/A. The peak looked reasonably close to what I'd imagine the TS to be, and I fed it to the dimer method using neb2dim.pl once more.

The dimer method still does not seem to be working as expected. The force goes down for the first few iterations but then continually increases. Why would this happen, and how can I address this kind of issue?

I apologize for making yet another post here -- this is the first time I've tried calculating a TS, and it's been an unsuccessful start. I appreciate anyone's advice. A .zip file containing the CI-NEB run and the follow-up dimer run is included for clarity.

[graeme]

The difficulty in this calculation is the fact that many atoms are moving along the reaction coordinate, but then a single atom is transferred at the transition state. This mean that you have a very long reaction path (sum of the motion of all atoms) which makes resolution along the path a challenge. In the attached calculations I used 8 images in the CINEB calculation, which nicely captures the saddle point. The attached dimer calculation also converges. Note that I used a very stable but somewhat slow optimizer (IOPT=3). My goal was to reach convergence stably rather than quickly.

[arosen]

Thank you very much for the helpful reply. I really appreciate it, as I've run into very similar issues with other TS I've searched for with this system.

I ended up very recently coming to the same conclusion you did - that the number of atoms that move between the initial and final state makes it difficult to sufficiently resolve the desired portion of the MEP unless many images are used. Using a higher number of images seems like the only way to go here (although I hadn't had success yet, so the workflow you used will be very helpful for me).

Interestingly, using the LBFGS optimizer with default options didn't seem to work very well in my CI-NEB attempts. The forces/energy would suddenly "spike" to very large values every ~20 iterations, followed by a return to the normal values after a few steps. This made convergence a bit of a challenge. I'll try QM like you used or FIRE if I continue to have trouble.

Thank you again for taking the time to look this over and providing such helpful feedback.

[arosen]

I should add that I think the CI-NEB run ended up capturing some sort of hopping mode of the radical instead of the H-abstraction barrier, unfortunately. That being said, the general idea of using a greater number of images (and a robust optimizer) will go a long way.