[Q-e-developers] Bug in meta-GGA interface with LIBXC

[Davide Ceresoli]

Dear Daniel,
     I had noticed exactly what you are reporting. I had
tested the QE implementation vs the libXC implementation
of metaGGAs and I've got a factor 1/2 in one case. The
problem is that I didn't know which one is correct!

Regarding the NaN's, I found that that the problem is the
ratio tau over tau_w (tau is the valence, kinetic energy, tauw
is the corresponding Thomas-Fermi). The correlation formula is
valid for z < 1, but I've noticed that inside the pseudo
radius, it happens that z > 1 over small regions of space.
Since I didn't want to "break" arbitrarily the TPSS implementation
in QE, I made the changes in my own repository (github=>ce-espresso)
and I've tried z = max(1.0,z) or using a smooth z = a*tanh(z/a)
formula. Other choices are possible, but yield different results.

I'm glad to know that now the problems are fixed! Thanks a lot.

Best wishes,
     Davide



On 11/11/2017 06:37 AM, Mejia Rodriguez, Daniel wrote:
> Dear QE developers,
> 
> 
> I'm using the PWSCF v.6.2 (svn rev. 13949:13950) interface with LibXC (version 
> 4.0.2) to perform some calculations with meta-GGA functionals. I noticed that 
> even for a very simple case, diamond computed with TPSS, QE interfaced with 
> LibXC yields NaNs and stops early; however, the native TPSS implementation runs 
> successfully.
> 
> 
> I started to look the source code for clues about this odd behavior and I 
> discovered an odd 1/2 factor in the kinetic energy density when LibXC is used.
> As far as I understand the code, that factor should not be present.
> 
> I kept digging into the code and found the following:
> 
> 1. The NaNs were coming from the correlation part of TPSS.
> 
> 2. By skipping the TPSS correlation call, I could manage to converge the SCF. 
> However, the results using only TPSS exchange for both LibXC and the native 
> implementation are different.
> 
> 2. Getting rid of the 1/2 factor in tau (and the corresponding 1/2 factor in 
> v3x) yields essentially the same result for LibXC and the native implementation 
> using only TPSS exchange.
> 
> 3. Getting rid of the 1/2 factor allows to use of LibXC TPSS correlation. 
> Furthermore, both the LibXC and the native implementation yield the same result.
> 
> 4. From my work with LibXC with other codes, I know that this interface can be 
> rather tricky. In order to further stabilize the PWSCF and LibXC interface I 
> added the same density threshold as in the native case, namely, skipping the 
> computation of the potential when the density gets below 1.0E-10 a.u.
> 
> 
> Note that the 1/2 factor appear in all calls to LibXC meta-GGA potentials, so it 
> should be corrected for all functionals.
> 
> 
> 
> I hope that my explanation is not too cumbersome. I'm also attaching the input 
> file I'm using (C.in) and the output files for the native (C.out) and the 
> corrected LibXC implementation (C.out.2).
> 
> If you have further questions, please do not hesitate to ask.
> 
> 
> 
> */Daniel Mejia Rodriguez/
> 
> Postdoctoral Associate
> QTP Department of Physics*
> University of Florida
> 2302 New Physics Building
> Gainesville, FL 32611
> (352)392-9227
> 
> 
> _______________________________________________
> Q-e-developers mailing list
> Q-e-developers at qe-forge.org
> http://qe-forge.org/mailman/listinfo/q-e-developers
>