UT theoretical chemistry code forum

Hi,

I have been running this diffusion calculation for over 1000 ionic steps now, and it refuses to converge. To give some history:

1) I guessed a pretty good path with 4 images, and forces were all around 1.
2) I ran with FIRE and after 100 iterations the forces were less than 0.5, and all looked good. Then the weird behaviour started, with forces jumping up every few iterations, and no sign of convergence.
3) After about 500 iterations just to see if I was being impatient, I went back to the 100 iteration point and switched on LBFGS.
4) This ran well for 100 iterations, and now forces were down to 0.2, but again it refused to converge and large jumps in the force started occurring.
5) I took a run with less than 0.2 forces, and added two more images between the longest parts of the path.
6) Force jumps are reduced, but this still does not converge after 200 iterations more…

So, do I just need more images? or is there an obvious remedy to this? I would really like an efficient way to calculate the barrier in this system, as this is just the control calculation before investigating atomic manipulation with a tip.

Cheers,

Adam

last run:

0 0.01434600 -857.06046900 0.00000000
1 0.19435400 -856.06434400 0.99612500
2 0.10395500 -855.18744600 1.87302300
3 1.77256500 -855.21995900 1.84051000
4 1.40782200 -855.30250200 1.75796700
5 0.70781700 -855.24955100 1.81091800
6 0.17177000 -856.12425200 0.93621700
7 0.01903000 -857.05822200 0.00224700

last run -2:

0 0.01434600 -857.06046900 0.00000000
1 0.14454600 -856.28187000 0.77859900
2 0.08409900 -855.53571700 1.52475200
3 0.13383500 -854.75648600 2.30398300
4 0.12392000 -855.20697000 1.85349900
5 0.05687200 -855.26419500 1.79627400
6 0.02271100 -856.18072900 0.87974000
7 0.01903000 -857.05822200 0.00224700

It all looks pretty reasonable, expect for LBFGS being unstable.

It does look like you have perhaps too few images. It is a long path with (what looks to me) at least 3 saddles. I can imagine an event in which a vacancy / adatom pair are formed, then diffusion of the adatom and diffusion of the vacancy, and then a final step in which the adatom fills in the vacancy again.

One strategy would be to minimize one of your high energy intermediate images (03 or 04) to see if there are intermediate minima along the path.
Calculating paths with a single saddle is faster and easier.

I would also suggest using as inexpensive calculation settings as possible for these calculations. Once you have found a saddle, there is little work involved in re-converging it with accuracy. For example, you can set PREC=normal, ENCUT = 280, and ISPIN=0 for the initial NEB runs.

If you want to try to converge using LBFGS with the long path, I suggest reducing INVCURV to 0.005 to improve stability.

Just to update this - after a further 100 steps with a couple more images, the path converges nicely...