[CP2K:5764] Energy Conservation Problems

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

Hi

yes, this looks like a force problem. Your diagnostics with the
non-correlation functionals is not fully conclusive (too short MD).
Still, it probably points in the right direction.

I would go to a small test system for conducting more basic tests.
For the tests I would not use any density smoothing. I would also
switch off the rVV10.

Test 1: Is the SCF convergence smooth and without problems. Use
        OT with Conjugated Gradients and line search to do this test.
        This gives you information on the validity of the functional
        derivatives.

Test 2: Use the DEBUG option to compare numerical and analytic forces.
        
You can also perform the same calculations using the standard B97
functional implemented in CP2K to get a feeling on the expected results.

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: cp... at googlegroups.com
From: August Melcher 
Sent by: cp... at googlegroups.com
Date: 10/23/2014 12:51AM
Subject: [CP2K:5764] Energy Conservation Problems

Hello all,
I've recently been working on implementing B97M-V, a new functional from Narbe Mardirossian in the Martin Head-Gordon group, in LIBXC for use with CP2K. I have in fact done so and been able to reproduce the Q-Chem binding energies on the S22 dataset up to a threshold we have deemed to be reasonable based on sample calculations with PBE-D3 and rVV10. However, I am running into trouble when moving to the condensed phase. I have attempted to simulate a 64 water molecule cluster with PBC but I see large amounts of drift in the energy. The attached plots show this drift. The first shows the drift over several ps, the second shows what the behavior looks like initially, over ~150 fs, and the last shows that oddly if I run without the local correlation term, with just exchange and the nonlocal rVV10 correlation, energy appears to be conserved. A sample input file is attached (BLYP pseudopotentials were found to give optimal performance on S22 for B97M-V).
In attempting to rectify this problem, I have tried (without success) the following:-Increasing the PW cutoff to 5000 Ry-Using the QZV3P basis set-Trying a different initial condition-Running without ASPC, using other WF guess methods including previous rho, previous WF, PS, and linear PS-Using a 0.25 fs timestep instead of 0.5 fs-Using an SCF convergence threshold of 10-8 instead of 10-7-Using the following smoothing/derivatives combinations: NN50/PW, NN50/SPLINE3, NN4/NN4, NN10/NN10
Does anyone have any idea what may be going on here? I only made changes to the LIBXC code, so perhaps I neglected to update some code which calculates the density derivatives in CP2K. Also the list of smoothing/derivative combinations is by no means exhaustive, so perhaps I just haven't found the proper combination for the numerics of this functional.
Thanks,August



-- 

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 http://groups.google.com/group/cp2k.

For more options, visit https://groups.google.com/d/optout.



[attachment "cons1.png" removed by Jürg Hutter/at/UZH]
[attachment "cons2.png" removed by Jürg Hutter/at/UZH]
[attachment "cons3.png" removed by Jürg Hutter/at/UZH]
[attachment "wb97xvwater.inp" removed by Jürg Hutter/at/UZH]