[CP2K-user] [CP2K:11230] Cell optimization of layered material, convergence vs K_POINTS

[hut... at chem.uzh.ch]

Hi

just a few suggestions that might help:

1) use EPS_DEFAULT 1.e-12 for additional numerical stability
2) use &REF_CELL with a cell of about the size of the largest cell expected
   this improves numerical stability by enforcing a constant number of grid points
3) use OPTIMIZER LBFGS
   I think it is more stable

best regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut... at chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp... at googlegroups.com wrote: -----
To: "cp2k" <cp... at googlegroups.com>
From: "Daniele Ongari" 
Sent by: cp... at googlegroups.com
Date: 02/04/2019 12:00PM
Subject: [CP2K:11230] Cell optimization of layered material, convergence vs K_POINTS

Dear CP2K developers, 

while optimizing the cell of 2d layered covalent-organic frameworks (think: graphite with holes) we are encountering instabilities that can seemingly be quenched by increasing the number of layers used in the cell (increasing the K_POINTS works as well).
Since the materials tend to have a PBE gap above 2eV and should be van-der-Waals bonded only, we find that surprising.

In the following we report one of these problematic cases, COF-1, using PBE-D3(BJ) functional.
Figure 1 shows the starting unit cell, which contains 2 layers of the material (before and after the cell_opt with kpoints).



Figure 2 shows the convergence of the total energy, once using just the Gamma point and once using 1x1x2 k-points (2 k-points along the z direction). Note that for the second one we use DIAGONALIZATION and EXTRAPOLATION=USE_GUESS, as required for using kpoints, but the same CELL_OPT settings.



The cell find its optimal dimension by tilting in the xy plane and, in the case of Gamma calculation, the optimization oscillates without converging.

- ot-gamma inp: https://www.dropbox.com/s/g5usnojtblhomi3/ot_inp.txt?dl=0
- kpoints inp: https://www.dropbox.com/s/d69170jnqfuq4g6/kp_inp.txt?dl=0
- coord+inputs+outputs zipped: https://www.dropbox.com/s/k3kl6sw9u8ine2y/cof1_ot_vs_kpoints.zip?dl=1

We used K-POINTS because the perpendicular width in z direction is 6.920 Angs at the start and gets 6.766 Angsat the end, however we are confused on why the K-POINTS in z directions are so important for the stability of the calculation. Is there any other parameter that I can tune to stabilize the no-K-POINTS+OT calculation?

I tested already many options without kpoints, without success (always experiencing jumps in the energy during the CELL_OPT):

1) lower EPS_SCF
2) higher CUTOFF
3) DIAGONALIZATION
4) EXTRAPOLATION=USE_GUESS

and with (a partial) success:

5) Duplicating the cell in z direction gives almost exactly the same result as with K-POINTS (and is more efficients in terms of speed, albeit not an elegant solution)
6) CG or (much) lower TRUST_RADIUS=0.1, do not jump but they are *very* inefficient, and therefore not a solution.
7) Using the PW code Quantum Espresso (much slower than CP2K for this system) with Gamma approximation. In that case PBE-D3 does always lead to a smooth cell_opt convergence using bfgs. As expected, the z dimension of the cell is small, but "not so small" to cause problems.

Therefore, the most efficient solution for now it would be to duplicate my 2D materials in the z direction (mod #5) to have a robust convergence with the price of a ~2x slower calculation and having a bigger reference cell. This does not *always* solve the cell_opt convergence problems, but it works for most of the problematic cases. Consider that, as in my previous post on OT convergence problems (https://groups.google.com/forum/#!msg/cp2k/X8RhYjSVVWc/TXdeLTMUCQAJ) our aim is to use general setting for high-throughput calculation. 

Do you have any other suggestion to try solving my problem?
I expect that these 2D-layers have a corrugated potential surface: I'm not looking for the absolute minima, but at least a reasonable one without geometry convergence problem. I'm wondering if there are some threshold value that I can tune not to have interference with the CELL_OPT, without increasing much the computational cost (not more than ~2x).

Any suggestion is welcome, thanks!

Daniele Ongari

  
  -- 
 You received this message because you are subscribed to the Google Groups "cp2k" group.
 To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns... at googlegroups.com.
 To post to this group, send email to cp... at googlegroups.com.
 Visit this group at https://groups.google.com/group/cp2k.
 For more options, visit https://groups.google.com/d/optout.