Open Andrew-S-Rosen opened 11 months ago
This is a known issue for the GFN1-xTB implementation. The xTB PBC support in xtb
should be deprecated now since there is a proper implementation via tblite
available.
Ah, okay thanks for the update. I feel like I did know at one point, but it must have slipped my mind.
After updating to xtb
latest version (6.6.1), I observed the same problem with the system cell (almost 2 times larger), even using the --tblite option:
xtb --tblite --chrg 0 --uhf 0 --gfn 1 --parallel 24 --opt tight test.POSCAR
The xTB PBC support in xtb
will not be maintained anymore?
Because I wanted to use the spin-polarization implemented in the latest version of xtb
for a periodic system, a feature that (I suppose) is not available in tblite
.
@alexamist Whike this doesn't help with the xtb executable, it turns out the spin polarized features are in tblite --- just not the ASE calculator. See https://github.com/tblite/tblite/issues/126.
@awvwgk: Can you confirm if PBCs in xtb
(the package) should not be used at all, or if it's okay to use if --tblite
(the flag) is set?
I know that tblite
(the package) is okay in terms of PBCs.
I can confirm that GFN1-xTB in xtb is still broken with PBCs and should be removed. I tested and verified GFN1-xTB from tblite mostly in dftb+. The parts provided by tblite in xtb should be fine, i.e. energy, forces and stress, not sure about the optimizer in xtb.
Thanks. That's my conclusion as well. Only thing I'm uncertain about is the --tblite flag with --opt in xTB. Will try to do a test.
Describe the bug When running
xtb POSCAR --opt --gfn 1
on the POSCAR below, the resulting converged structure has bond lengths that are roughly 2x as long as physically reasonable.When doing the geometry optimization with the
tblite
standalone package via ASE as the optimization wrapper, everything works fine. When runningxtb POSCAR --opt --gfn 1 --tblite
, everything works out as expected as well I believe.This is with the most recent
conda
builds of bothxtb
andtblite
on Ubuntu in an otherwise clean Conda environment. I haven't yet tried older versions ofxtb
, in case it's related to #884, but that's probably the thing worth trying next on my end.To Reproduce See above.
xtb POSCAR --opt --gfn 1
POSCAR.txt
Refer to the CONTCAR below, which was made from
xtbopt.vasp
. Note how the bonds are all extremely elongated.CONTCAR.txt
Expected behaviour The bonds should not be elongated to unphysical lengths (the unit cell here ends up being almost exactly 2x larger than expected).
Additional context Seems like something is suspect regarding the PBCs in standalone
xtb
, at least with GFN1-xTB.Tagging @prathami11 who initiated this report with me.