|
|
我只是邮件搬运工。。。
The main new features are
- multipole electrostatics
- removal of all classical corrections as well as special pair
potentials/parameters,
. i.e., its less empirical
- ATM dispersion term
- atom specific CN dependent level shift and one-center XC-anisotropy term
- much better parameterization (more reference systems, better frequency
fit)
What is still missing (but likely will have only a small effect at this
accuracy level)
is the D4 instead of D3 dispersion model which will be implemented
soon. We are also working on the GFN2 paper but this may take some weeks.
Many properties and problems have been improved or solved compared
to GFN1. Most notably are non-covalent interactions and conformational
energies improved
and most metals are described more consistently. However, there are
of course also drawbacks and on the downside we note sometimes less
stable SCF
convergence behavior. Both, this and the higher accuracy is related to
the short range part of the
multipole electrostatics. In non-convergent-SCF cases GFN1 is still
available as fall-back option.
The new method is on average only insignificantly slower than GFN1 (even
faster for systems with many hydrogens) and we think that GFN2-xTB
offers an extremely good cost/performance ratio often reaching
DFT-D3 accuracy for the target properties. All non-covalent interactions
are now exceptionally good including water, and halogen or pnicogen
interactions.
Additionally some typical thermochemistry properties are described better
(without fitting them!) which will be discussed in detail in the
forthcoming paper.
In summary I don't see much room for further general improvement at this
high-speed QM level.
To run GFN2 just copy the parameter file .param_gfn2.xtb
to the xtb path and set the default in the .xtbrc file
(line two). GFN1 is still available by the flag -gfn1
(or setting the default of variable gfnver to 1).
Important note: the parameters are preliminary, i.e.,
it make little sense to start big projects @GFN2 until
we got feedback and may be change a few details
(which is more likely to happen for heavier, "uncommon" elements
than for HBCNOFCl.. for which the present parameters
are probably the final ones). In addition we're right now
re-adjusting the GBSA parameters to GFN2. The present GFN1
GBSA parameters can be used, however, without
much loss of accuracy.
Please report any problems, bugs, or bad cases
but we're also thankful for success stories
(Christoph Bannwarth and I invested some time into all this and
if was often but not always fun ...).
Those of you who sent me GFN1 outliers (thanks again!)
may check if they are now qualitatively ok with GFN2 or
at least have improved.
|
|
发表于 Post on 2018-3-31 18:18:02
收藏 Add to favorites
发表于 Post on 2018-4-10 22:24:28
楼主 Author