Dear All,
I've been experiencing some accuracy issues while doing (both ground state and
response-function) calculations with ABINIT (I'm using version 6.12.3) and
Troullier-Martins (FHI, type 6) pseudopotentials with non-linear core correction
(NLCC). The problem is that there is a large net force (violation of the acoustic
sum rule) on the atoms of the cell. This issue was hard to spot at first, as ABINIT
does an automatic "fix" of the ASR on the Cartesian forces -- one has to look
at the dE/dt information. This net force _does not_ go away by simply
increasing the energy cut-off.
To give you a flavor of the problem, I have performed some cross-checks with a
different code, LAUTREC, where I can use the same identical pseudopotentials,
XC functional (ixc=7, LDA PW92) and basis set. Indeed, when not using NLCC,
the two codes are in excellent numerical agreement (typically 10^-6 Ha in
the total energy values). When atomic species using NLCC are present, the agreement
in the total energy worsens somewhat (to 10^-4 Ha). This is not yet a reason to worry in
principle, as different codes do somewhat different things. However, if I compute the
total forces along y on a sample system that I shall explain below, I obtain the following
results:
------------------ 25Ha --------- 35Ha
ABINIT 0.000004001827 0.000001075583
LAUTREC 0.000000002215 -0.000000000045
The net drift is _large_ in ABINIT, and does not appear to decrease that much with the cut-off,
unlike the LAUTREC results. At 35Ry, the drift is of the order of 4*10^-10 Ha/bohr in LAUTREC, 1*10^-6
in ABINIT, i.e. 3-4 orders of magnitude difference!
The system is a 4-atom NaCl cell, where I have displaced one Na atom by a small distance
in the yz plane (along 110). You can find the complete files necessary to run the ABINIT
calculation at the following URL:
http://www.icmab.es/leem/stengel/nacl_test.tar
Note that I run into the same issue, if not worse, when attempting response
functions (phonon) calculations. I see two possibilities to explain this fact:
1) I overlooked some important parameter in my input files (or in the header
of the na.fhi file, that I prepared myself), which would explain the lack
of accuracy in the atomic forces.
2) The implementation of the non-linear core corrections are simply different
in LAUTREC and ABINIT. If this turns out to be the case, I think it would be
maybe worth checking if there's anything that can be done in the code to
improve the accuracy of the NLCC part.
Thanks in advance for helping me out with this.
Best regards,
Massimiliano