[CP2K-user] [CP2K:15690] Re: SCF convergence problem with large basis sets

hut... at chem.uzh.ch hut... at chem.uzh.ch
Tue Jul 6 08:04:27 UTC 2021


in GAPW calculations the electron density on the regular grids depends 
on the soft/hard basis separation. This value will converge to some
irregular value, but it is impossible to predict it.
More important is the total electron density. If it is off from the
integer number of electrons, you see errors from two sources.
1) the cutoff from the regular grids (as in GPW)
2) the atomic size approximation 
Errors from 1) can be reduced by increasing the cutoff, errors from
2) are related with the rcut value of the ALL_ELECTRON file.
"Carefully" adjusting that value can improve high precision results.


Juerg Hutter
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: "ma... at gmail.com" 
Sent by: cp... at googlegroups.com
Date: 07/06/2021 01:56AM
Subject: [CP2K:15690] Re: SCF convergence problem with large basis sets

Hi Nick,

Just curious about your third suggestion (the CUTOFF issue). 
When I use ALL_ELECTRON basis sets, I found increasing CUTOFF does not change the second value of "Electronic density on regular grids" at all, although the SCF converged energy is decreasing. I tried DZVP all electron basis set for Si. But even using a CUTOFF as high as 1000 Ry, the second value of "Electronic density on regular grids" is still 13.5173370396 same as CUTOFF of 50 Ry. I'm wondering do I need to worry about this when I use ALL_ELECTRON basis sets?


在2021年7月6日星期二 UTC+10 上午8:49:27<n... at berkeley.edu> 写道:


(1) When you say that you had no issues with  DZVP -> TZVP -> TZV2P basis sets, were you using the MOLOPT basis sets? The MOLOPT basis sets were optimized using the overlap matrix condition number as a constraint in order to make them more numerically stable. This is why they are the basis set type of choice for condensed phases. If you *were* using MOLOPT, then once you move to the QZVP basis sets, for which there are no molopt basis sets, then that is why it become harder to converge. If you *were not* using the MOLOPT basis sets I would encourage you to stick with them. Generally, condensed matter systems have converged properties around TZVP quality in my opinion, and DZVP is still pretty good. If you need true chemical accuracy, then you're going to need to move beyond DFT anyway.

(2) There is a general issue of using larger basis sets, which is the nature of the Gaussian type orbitals. GTOs are not an orthonormal basis, unfortunately, so the larger your basis set, the greater the risk of introducing linear dependencies that make converge very difficult. Another reason to limit the size to only as large as you need for your application. Beyond linear dependencies, the conditioner number itself also increases with increasing basis set size.

(3) Your CUTOFF in your multi-grid is not converged. I noticed this because after your SCF loops you have the line "Electronic density on regular grids. -605.9997903253        0.0002096747" -- the second number in this column should be <1e-8 preferably. The cutoff is the most common cause for this, and your cutoff of 350 is not sufficient. To determine the cutoff, take the largest exponent in your basis set and multiply it by the relative cutoff. Your CUTOFF value should be at least this large, otherwise your multigrid will not be able to accommodate the hardest exponents. Oxygen has an exponent of ~12 at the QZV3P level of theory, so your CUTOFF should be around 480 if you are using the default REL_CUTOFF of 40. The exponents change with your basis set, so this could be part of the issue you were facing when you got to the larger ones.

Try fixing your CUTOFF value and see if it helps, but also consider using smaller basis sets, maybe of the MOLOPT type, which are generally sufficiently accurate for most DFT calculations. 

On Monday, July 5, 2021 at 5:02:47 AM UTC-7 tom... at ugent.be wrote:
Sorry forgot the attachments.

On Monday, July 5, 2021 at 1:44:50 PM UTC+2 Tom Braeckevelt wrote:
Dear CP2K users/developers,

I was performing some benchmarks related to the basis set size (with simply PBE-D3). In particular, I was checking the energy difference between benzene adsorbed in the H-SSZ-13 zeolite and its protonated counterpart. While I can see a steady improvement going from DZVP -> TZVP -> TZV2P (without any issue in the calculation) I encountered large problems in the SCF convergence when using even larger basis sets, like the QZV2(or 3)P and the augmented basis sets. At this point, I was only able to converge a couple calculations with the QZV3P using the CG optimizer (normally I use the DIIS) and the FULL_KINETIC preconditioner (I normally use the FULL_SINGLE_INVERSE), which according to the manual should be more robust. I also tried the diagonalization method instead of the OT one but also in that case the scf was not converging.
The problem is that even with the two converged calculations the energy difference between the two species I'm studying results to be ~3000kJ/mol vs the expected ~100kJ/mol, thus it looks like the two scf have converged to different minima.

Does anybody know how to robustly converge the scf with large basis sets? In attachment the input and output of the supposedly converged calculations.

Thank you in advance,
 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 cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/50b279f0-7e3c-4e39-a80f-379cd3f651d1n%40googlegroups.com.

More information about the CP2K-user mailing list