Bulk potential energy by different methods

Matt W mattwa... at gmail.com
Fri Jun 2 13:55:23 UTC 2017

On Friday, June 2, 2017 at 2:43:23 PM UTC+1, Sebastian Hütter wrote:
> Hi,
> well, that would be far more than 1 vacancy per "infinite" volume. For 
> these I use a coordinate list corresponding to 5x5x5 unit cells (that is, 
> 500 atoms in 20.25 angstroms³), compute the energy (E1), remove one atom, 
> compute again (E2), then calculate Evf = E2 - (499/500) * E1.
> With the exact same parameters, all EAM potentials I compared are already 
> well converged (i.e. larger replications change less than 1% of the 
> absolute value, so we're out of the self-interaction range), and get within 
> 30% of the experimental value (~0.66eV) (depending on their construction of 
> course).
> Different methods converge at different rates. 

> Reducing EPS_DEFAULT helped in that regard and I now get somewhat 
> plausible results there. However that never really was my objective, I just 
> chose vacancy formation energy to test because I it was the first concrete 
> physical quantity that came to mind with a simple computation method.
> What I was really looking for is a way to compare the energies over things 
> like structural changes, but that doesn't appear to be possible... I do 
> note that every energy I have compared so far has been different by the 
> same constant factor (15.8). If there was any physical reason for that, 
> there's an obvious solution to my problem, but I can't think of one.
> Is should be - cohesive energies, structural relaxations do not use a 
'total energy' but you subtract one or more total energies and work with a 
difference. That should be well defined and comparable between methods (and 
to experiment).


> I've since found some articles that compared DFT to EAM for PdAu 
> <https://journals.aps.org/prb/abstract/10.1103/PhysRevB.80.035404> and 
> NiAl <http://iopscience.iop.org/article/10.1088/0965-0393/24/4/045012> 
> systems and where it worked well (both with VASP though).
> Checking with the JARVIS-FF 
> <http://www.ctcms.nist.gov/~knc6/periodic.html> repository shows a 
> cohesive energy of on average -3.37eV/atom and which was confirmed by DFT, 
> so I am now under the impression that the method should work and something 
> in my attempt to do the same with cp2k is wrong.
> Sebastian
> Am Freitag, 2. Juni 2017 12:45:15 UTC+2 schrieb Matt W:
>> How do you do the vacancy calculation? Are you still using multiple unit 
>> cell and just removing an atom?
>> Matt
>> On Friday, June 2, 2017 at 11:25:48 AM UTC+1, Sebastian Hütter wrote:
>>> Hi Matt, Matthias,
>>> thank you for your replies.
>>> The input file is basically identical to a similar question about Nickel 
>>> posted here some time ago 
>>> <https://groups.google.com/d/msg/cp2k/llUH48_Y3vY/GPrPXaaXZJIJ>. I did 
>>> convergence studies on EPS_SCF and cell replication count NREP; the 
>>> parameters posted are the fastest in terms of simulation time. For the Evf 
>>> calculations, I used 5x5x5 unit cells (500/499 atoms) for both methods.
>>> individual total energy are almost always meaningless, especially when 
>>> comparing different levels of theory.
>>> I'm still new to this, but... why? Shouldn't the total energy be a 
>>> direct consequence of the cohesive energy, especially in such simple 
>>> systems?
>>> a gamma point calculation for Al metal using a model system consisting 
>>> of only 3x3x3 unit cells won't provide converged results. You need a much 
>>> larger model system in the order of 10x10x10 unit cells or you have to 
>>> employ an equivalent k point sampling. Besides that major issue, there are 
>>> further shortcomings in your CP2K input: (1) an SCF convergence threshold 
>>> of 1.0E-4 (EPS_SCF) is too large: 1.0E-6 (or smaller), (2) a general 
>>> threshold of 1.0E-10 (EPS_DEFAULT) is also too large: 1.0E-12 (or smaller)
>>> I have redone the calculation with your suggested thresholds and 6x6x6 
>>> unit cells which is about the largest I can run on our local compute server 
>>> because of memory constraints, if that showed any change at all I would 
>>> have to post to the HPC cluster. However: no change. Still -8.38Ha/UC.
>>> Is LDA a good enough density functional. Is a DZVP basis set good enough?
>>> I don't know? How do I find out?
>>> Did you relax the structures?
>>> A CELL_OPT run changed the total energy by just a few percent (for this 
>>> setup, a=4.05 is a bit on the high side), not the order of magnitude that's 
>>> missing here.
>>> Kind regards,
>>> Sebastian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20170602/4f6fb113/attachment.htm>

More information about the CP2K-user mailing list