trouble in converge

[yetchen]

Dear Prof. Henkelman,

We try to study a C-S bond-breaking process and uploads three files for the same process: 01, 02 and 03.
For 01, we set ediffg = -0.1 with iopt = 3, and the job is successfully converged.
For 02, starting from the 01, we further set ediffg = -0.05 with iopt =3, however the job seems to be unconverged (the uploaded file is the results at 94 steps). The force at the saddle point remains around -0.1.
For 03, based on 01, we set ediffg = -0.05, iopt =1 and ediff = 1e-7. The force grows even higher, from 1 to over 10.

Based on 01 results, are there any good ways to converge the force when we set ediffg = -0.05?

Thank you all.

[graeme]

The calculations in 01 and 02 look ok. In 02, the force on the saddle does remain around 0.1 eV/Ang, but the energy is dropping systematically. Note that the barrier has dropped by over 0.1 eV. I would continue the calculation from 02 using the same optimizer. If you want to switch to the LBFGS optimizer, you could try lowering the INVCURV value.

More importantly, I notice that you have no frozen atom in your substrate. This is typically done to model the bulk material. It also helps to stabilize the optimizers by avoiding low modes and translation.

[yetchen]

Dear Prof. Henkelman,

Thank you for your suggestions, and we will do these tests.
Best wishes!

[yetchen]

Dear Prof. Henkelman,

We encounter the same problem when dealing with two different systems. The job files are uploaded as the System1 and System2. In both of these jobs, the 01-05 images are generally 2~5 eV over the initial state (IS, 00) and final state (FS, 06). We suspect that some kind of difference in INCAR setting when optimizing the IS (FS) and NEB, leads to the problem. So the question is how can we avoid this unusual shift-up of energy?

Thank you.

[graeme]

It looks like your endpoints were spin polarized.

[yetchen]

Dear Prof. Henkelman,

We try two different INCAR settings for frequency convergence, but the results seems that the parameters are not suitable.There’s two questions confusing us. The job files are uploaded as the freq1 and freq2

1. Are the two INCAR files correct ?
2. We set ‘ NSW = 1’, while using first INCAR the steps come up to more than 170, and the steps are 49 when using second INCAR.Where is the problem come from?

Thank you.

[graeme]

First, this question does not appear to be related to our vtstcode. But anyway, the calculation will need to do 3N+1 force calculations to get the dynamical matrix, where N is the number of free atoms in the system. If you have 150 atoms, you will need 451 force evaluations.

[yetchen]

Dear Prof. Henkelman,
We encounter some problems using the force based optimizer ‘IOPT =7’ .The job files are uploaded as the TS4.

1. After 53 steps compulations, the force of 03 image suddenly rises to 40. Is this a normal condition?

2. The speed of compulation seems slow for the coarse parameter settings. Is there any parameter setting can improve the speed?

Thank you !

[graeme]

It looks like image 03 found a new low-energy structure and the forces became high which destabilized the optimizer. You can decrease the TIMESTEP variable to deal with that. But first, I suggest minimizing image 03 separately and you might find an intermediate minimum, or new final state for your path.

[yetchen]

Dear Prof. Henkelman,

We set the EDIFFG= -0.1 for a coarse convergence, but the speed of the calculation is too slow to carry on. The job files are uploaded as the
TS. Is there any parameter settings you suggest to change?

Thank you!

[graeme]

I have a couple of comments about this system.

First, with DFT+U, there can be problems with convergence of the electronic structure. If the calculation hits the maximum number of electronic iterations, it can give high forces as a result of the unconverged electronic structure. You might try other IALGO tags, such as damped dynamics, which tends to be more stable for tough systems.

Second, your reaction has water dissociation at an oxygen vacancy site on ceria two form two adsorbed hydroxyl. In your initial path, however, the dissociated H has to transfer all the way across a surface Ce atom to reach the specified O adsorption site. There is a closer oxygen atom which makes more sense for accepting the H. In the attached pathway, you can see that this leads to a much shorter pathway. As I mentioned previously, the long path that you choose is probably not an elementary reaction and there would be an intermediate minimum (e.g. H2O dissociation and then H diffusion). More images would likely be required to resolve that path between your initial and final states.

Third, and I can't say this enough times, use inexpensive settings for calculating reaction pathways and any exploration of the energy landscape. Once you know the relevant paths and reaction mechanisms, it is easy to reconverge the calculation with more accurate settings. In the attached, for example, I've switched to a gamma point calculation and a minimal cutoff of 300 eV. These settings will make the calculation about an order of magnitude faster.

Finally, you can take a look at the attached NEB calculation which shows a near-barrierless dissociative adsorption process. [Note that I switched the reactant and product.] It is not fully converged, but it already shows the mechanism and barrier. It is also converging just fine.