[CP2K-user] [CP2K:15581] Poor force convergence wrt basis set size

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

Hi

If linear dependencies are a problem you can try
&SCF
   CHOLESKY OFF
   EPS_EIGVAL  xxx
&END
but this will not work with OT.

There is also a new (not tested) set of basis functions that
combine the cc polariaztion and augmentation functions with
the PP valence set. 
see BASIS_ccGRB_UZH in the latest TRUNK version.

regards

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: cp... at googlegroups.com
From: "Arthur France-Lanord" 
Sent by: cp... at googlegroups.com
Date: 06/16/2021 05:33PM
Subject: Re: [CP2K:15581] Poor force convergence wrt basis set size

That sounds reasonable. It's interesting to see that DZVP-MOLOPT-SR actually performs quite well when compared to QZV3P ; on par with QZV2P. 

I've been trying to use a large, generic basis set such as one from the aug-cc family, but the calculation always fails with the usual Cholesky decomposition failed message. I've tried different preconditioners, which didn't help. Do you have any tips for that? Is OT more sensitive to overcompleteness than other methods? 

best, 
Arthur
Le mer. 16 juin 2021 à 13:35, <hut... at chem.uzh.ch> a écrit :
Hi
 
 my guess (no proof available) would be that your specific system
 with only S-S bonds is at the edge of the MOLOPT optimisation space.
 The different variations (SR vs. orig) have been optimized for
 energy and might be comparably sub-optimal for your system.
 
 The set of S containing molecules used in the optimization is
 
 Al2S3
 CS2
 H2SO4
 S2
 S5
 SF2
 HSSH
 SF4
 SF6
 
 regards
 
 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: cp... at googlegroups.com
 From: "Arthur France-Lanord" 
 Sent by: cp... at googlegroups.com
 Date: 06/16/2021 11:47AM
 Subject: Re: [CP2K:15578] Poor force convergence wrt basis set size
 
 Hi, 
 
 Thank you for the suggestion. This contributed to reducing the noise on the forces, and MAE's are now ~ 30% smaller. However, I still observe almost similar performances for MOLOPT basis sets (from DZVP-MOLOPT-SR to TZV2P-MOLOPT, vs. TZV2PX-MOLOPT). In fact, the MAE is even slightly smaller for the smallest, short-range version of DZVP-MOLOPT. Could that be a symptom of over-completeness of basis sets including diffuse functions for this system? 
 
 best, 
 Arthur
 Le mer. 16 juin 2021 à 10:56, <hut... at chem.uzh.ch> a écrit :
 Hi
 
  I to be sure that you don't pick up basis set dependent numerical noise
  I would suggest to use a more strict eps_default value, e.g. 1.E-14.
 
  regards
 
  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: "Arthur France-Lanord" 
  Sent by: cp... at googlegroups.com
  Date: 06/06/2021 12:01PM
  Subject: [CP2K:15521] Poor force convergence wrt basis set size
 
  Hi, 
 
  I'm currently testing atomic force convergence with respect to basis set size, for a crystalline polymorph of sulfur (an array of cyclic S8 molecules). I've tested both standard GTH basis sets (DZVP to QZV2P compared to QZV3P), and MOLOPT (DZVP-MOLOPT-SR to TZV2P-MOLOPT compared to TZV2PX-MOLOPT). I'm observing a slow or poor convergence behavior, which can be seen in these two plots (MAE : mean absolute error, MAPE : mean absolute percentage error, both only computed on the x components of the forces): 
 
  GTH basis sets: 
 
 
  MOLOPT basis sets (the TZV2P-MOLOPT/QZV3P panel is just shown here for comparison) : 
 
 
  In particular, TZVP-MOLOPT performs slightly better than TZV2P-MOLOPT..? I'm using a high energy cutoff (2500 Ry and a REL_CUTOFF of 100 Ry), and a rather tight SCF convergence criterion (1.0e-7) with the OT method. Convergence in the SCF is achieved rather fast. 
 
  Are these errors expected? I was trying to reproduce something similar from what ccan be seen in https://arxiv.org/pdf/1409.4527.pdf (figures 1 and 2), but didn't expect such high relative errors. The configuration I'm using is from a molecular dynamics run to obtain non-zero forces. I'm including a cp2k input file to reproduce the QZV3P calculation. 
 
  many thanks, 
  Arthur
 
 
    -- 
   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/8d3ba5a1-9d0f-4dfe-8426-df09318745a9n%40googlegroups.com.
 
 
  [attachment "cp2k.inp" removed by Jürg Hutter/at/UZH]
  [attachment "geometry.xyz" removed by Jürg Hutter/at/UZH]
 
  -- 
  You received this message because you are subscribed to a topic in the Google Groups "cp2k" group.
  To unsubscribe from this topic, visit https://groups.google.com/d/topic/cp2k/-TXuRXyRtxs/unsubscribe.
  To unsubscribe from this group and all its topics, send an email to cp... at googlegroups.com.
  To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/OF98A1FDDF.1A7050A2-ONC12586F6.003111C8-C12586F6.003111CA%40lotus.uzh.ch.
 
   -- 
  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/CAAsXd9YYVN_0R%2BeUnFh2NddACOqC68dNLptwGoX0oGhNsu1K5A%40mail.gmail.com.
 
 
 -- 
 You received this message because you are subscribed to a topic in the Google Groups "cp2k" group.
 To unsubscribe from this topic, visit https://groups.google.com/d/topic/cp2k/-TXuRXyRtxs/unsubscribe.
 To unsubscribe from this group and all its topics, send an email to cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/OF20A92BBB.80A39778-ONC12586F6.003F2495-C12586F6.003FA2FE%40lotus.uzh.ch.
   
  -- 
 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/CAAsXd9a6qFdf%3D5nWUGcxLrLzHy9SneXciMDmNEa38Ec6grBcfA%40mail.gmail.com.