[CP2K:9469] Re: Hybrid functional calculation results show large difference with VASP

[S Ling]

Hi

Looks like you may changed your input from the initial one slightly: the
PDOS analysis shows you are running spin polarised calculations (at least
for CP2K, as the sum equals to 1; you didn't include UKS or LSD in your
initial input). For the analysis on VASP calculation, are these numbers
correspond to all the atoms in the cell or just part of it? The sum of the
numbers doesn't equal to 1 or 2, so we cannot see what are the real
differences.


"I will also try to do a ternary system like CsPbBr3."

>From your benchmark data, I would be more worried about In and Ag. I cannot
find any ternary halide involving Ag and In, but you can find AgInS2 from
ICSD database. I think this may be a better system to test. You may also
try CsAgX and CsInX.


"I have another question for Cs2InAgCl6. You are saying that Ag is
optimized for +1. What I am going to study is the hole state of Cs2InAgCl6,
so one electron will be removed from the system. In that case, Ag will tend
to be in 2+ state. Can you comment on the reliability of the PBE0
calculation if I focus on the structure instead of  the band gap?"

I think the uncontracted ADMM basis sets (FIT??, not cFIT??) should be
flexible enough to describe Ag2+. From your CP2K/PDOS analysis, it looks
like there may be a competition in terms of hole localisation, as the HOMO
features mixed Ag+ and Cl- states. I am not sure how the smaller band gap
in CP2K (compared with your reference) will affect hole localisation
in Cs2InAgCl6.
Again, I would start from a simpler binary system as a testbed, e.g. AgCl,
which I am sure has been studied before, I would use the same PBE0-TC
functional, and I would then compare the hole localisation properties (e.g.
self-trapping energy, etc) of AgCl with previous literatures. Once I am
happy with the results on binary system, I would then feel more comfortable
to move to the more complicated ternary and quaternary systems.

An additional test you can try is to relax the cell paramters and atomic
positions of Cs2InAgCl6 in CP2K at PBE0-TC level, and see if that improves
the agreement with the experimental value on band gap. I was looking at a
JPCL paper (DOI: 10.1021/acs.jpclett.6b02682) on Cs2InAgCl6, and it looks
like the authors reported a range (2.9~3.3 eV) of band gap values at PBE0
level for Cs2InAgCl6 with different geometry relaxations. This actually
shows the band gap is quite sensitive to small differences in geometries.

It looks like this discussion has become quite technical now. I am happy to
continue the discussion offline if you are interested.

SL



On 21 September 2017 at 18:48, Xiaoming Wang <wxia... at gmail.com> wrote:

> Hi,
>
> Here is PBE0 PDOS of HOMO and LUMO of Cs2InAgCl6.
>
> CP2K
>
> VASP
>
> s
>
> p
>
> d
>
> s
>
> p
>
> d
>
> HOMO
>
> Cs
>
> 0.000
>
> 0.000
>
> 0.000
>
> 0.000
>
> 0.000
>
> 0.000
>
> In
>
> 0.000
>
> 0.000
>
> 0.050
>
> 0.000
>
> 0.000
>
> 0.030
>
> Ag
>
> 0.000
>
> 0.000
>
> 0.454
>
> 0.000
>
> 0.000
>
> 0.368
>
> Cl
>
> 0.003
>
> 0.488
>
> 0.004
>
> 0.000
>
> 0.456
>
> 0.000
>
> LUMO
>
> Cs
>
> 0.085
>
> 0.000
>
> 0.000
>
> 0.000
>
> 0.000
>
> 0.000
>
> In
>
> 0.300
>
> 0.000
>
> 0.000
>
> 0.332
>
> 0.000
>
> 0.000
>
> Ag
>
> 0.241
>
> 0.000
>
> 0.000
>
> 0.212
>
> 0.000
>
> 0.000
>
> Cl
>
> 0.194
>
> 0.074
>
> 0.107
>
> 0.096
>
> 0.048
>
> 0.024
>
> I will also try to do a ternary system like CsPbBr3.
>
> I have another question for Cs2InAgCl6. You are saying that Ag is
> optimized for +1. What I am going to study is the hole state of Cs2InAgCl6,
> so one electron will be removed from the system. In that case, Ag will tend
> to be in 2+ state. Can you comment on the reliability of the PBE0
> calculation if I focus on the structure instead of  the band gap?
>
> Best,
>
>
> On Thursday, September 21, 2017 at 11:10:15 AM UTC-4, S Ling wrote:
>>
>> Hi
>>
>> I am not sure how you ran the PBE0 calculation without ADMM using the
>> MOLOPT basis sets, which were not designed for hybrid DFT calculations.
>> Hybrid DFT calculations without ADMM for your Cs2InAgCl6 system will use a
>> lot of memory, and as Prof Hutter suggested earlier, it will also be
>> 100~1000 times slower.
>>
>> "without ADMM basis sets I have SCF convergence issue if I don't use the
>> Coulomb truncation method, both for OT and diagonalization, the SCF didn't
>> tend to converge. Why is that?"
>>
>> Please have a look at the two papers below, which answer your question:
>>
>> http://pubs.acs.org/doi/abs/10.1021/ct900494g
>> http://aip.scitation.org/doi/10.1063/1.2931945
>>
>> For the PBE0 band gaps calculated using different codes, I can see for
>> binary systems, the differences are at the order of 0.2 eV or smaller,
>> which I think is not something totally unexpected, because of the different
>> HFX implementation strategies, different basis sets and different
>> pseudopotentials used by the two codes. If you try a third code based on
>> atomic orbital basis sets, you may also get different answers. In addition,
>> similar results at PBE level doesn't necessarily mean the numbers at PBE0
>> level will or should be similar as well. From computational point of
>> view, compounds involving d block elements are usually more difficult to
>> deal with than compounds involving only s and p block elements. As you
>> mentioned earlier, it could simply be the case that the differences in
>> binary systems added up in your quaternary system. It may be useful to
>> check what numbers the two codes will give for the ternary system, before
>> we look at the more complicated quaternary system. The VASP/PROCAR and
>> CP2K/PDOS analysis will tell us what are the nature of the states at the
>> top of the valence band and at the bottom of the conduction band, which
>> will help us to understand the difference between the two codes.
>>
>> SL
>>
>>
>> On 21 September 2017 at 14:57, Xiaoming Wang <wxi... at gmail.com> wrote:
>>
>>> I'd like to add that my PBE0 calculation without ADMM basis sets has
>>> been done. I tested 10*10*10 Cs2InAgCl6, which is the system of my original
>>> post. The band gap is 2.52 and 2.50 eV for with and without ADMM. So it
>>> seems not the ADMM basis problem?
>>> I have another question that without ADMM basis sets I have SCF
>>> convergence issue if I don't use the Coulomb truncation method, both for OT
>>> and diagonalization, the SCF didn't tend to converge. Why is that? (The
>>> calculation is fine with TC.)
>>>
>>> Best,
>>>
>>>
>>> On Thursday, September 21, 2017 at 8:54:49 AM UTC-4, Xiaoming Wang wrote:
>>>>
>>>> Hi,
>>>>
>>>> I am sorry for the confusing formula. It is actually Cs2InAgCl6, in
>>>> which Ag is +1. When you are fitting the ADMM basis sets, what's the
>>>> reference system for Ag?  Is it in solid environment? Is it possible for me
>>>> to fit the ADMM basis myself based on my vasp calculations? Well, maybe it
>>>> is too difficult for me, since I am new to cp2k. Btw, do the ADMM basis
>>>> sets depend on the oxidation states of the elements involved? If that's the
>>>> case, there would be problems dealing with different systems with the same
>>>> basis sets. Moreover, the oxidation state is an arbitrary quantity
>>>> depending on the charge partition scheme. The oxidation state may be slight
>>>> different even for the systems which are assumed to be +1 for a particular
>>>> element, for example.
>>>>
>>>> Best,
>>>>
>>>> Xiaoming
>>>>
>>>> On Thursday, September 21, 2017 at 6:17:45 AM UTC-4, S Ling wrote:
>>>>>
>>>>> Hi
>>>>>
>>>>> Can you confirm the chemical formula of your CsInAgCl system? I looked
>>>>> through the ICSD database, and the only related compound which I can find
>>>>> is CsAgInF6, in which Ag is in an uncommon +2 oxidation state (in your
>>>>> benchmark test, you looked at AgCl, in which Ag is in the common +1
>>>>> oxidation state). If this is the case for CsInAgCl6, there is a possibility
>>>>> that your CP2K/PBE0 and VASP/PBE0 calculations may have converged to
>>>>> slightly different SCF solutions. You can check this by comparing the
>>>>> orbital occupation numbers from your VASP/PROCAR and from the CP2K/PDOS
>>>>> analysis. Another possibility is that the ADMM basis sets are indeed not
>>>>> good enough to describe Ag2+ (I have only considered Ag+ when I was fitting
>>>>> the ADMM basis sets of Ag).
>>>>>
>>>>> SL
>>>>>
>>>>>
>>>>>
>>>>> On 21 September 2017 at 03:52, Xiaoming Wang <wxi... at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> Hi, Can you see it this time?
>>>>>>
>>>>>>
>>>>>> <https://lh3.googleusercontent.com/-f1GOmo2FRic/WcMpJhTURnI/AAAAAAAAGcU/MqO1xDMjsmQIV1PsgFDq5Aq_IqaPdz-CQCLcBGAs/s1600/benchmark.PNG>
>>>>>>
>>>>>>
>>>>>> On Wednesday, September 20, 2017 at 10:45:10 PM UTC-4, Matt W wrote:
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> I only see the check for InCl3, not the main system?
>>>>>>>
>>>>>>> Matt
>>>>>>>
>>>>>>> On Thursday, September 21, 2017 at 10:36:15 AM UTC+8, Xiaoming Wang
>>>>>>> wrote:
>>>>>>>>
>>>>>>>> Hi Matt,
>>>>>>>>
>>>>>>>> Actually I also did the calculation for a supercell of 21*21*21
>>>>>>>> with 320 atoms. PBE0 without TC are also tested, please see my benchmark
>>>>>>>> tests in previous post (attachment in reply to Ling).
>>>>>>>>
>>>>>>>> Best,
>>>>>>>> Xiaoming
>>>>>>>>
>>>>>>>>
>>>>>>>> On Wednesday, September 20, 2017 at 10:27:48 PM UTC-4, Matt W wrote:
>>>>>>>>>
>>>>>>>>> That ADMM basis is already pretty large, so whilst it might be the
>>>>>>>>> problem, first I'd check the cell size.
>>>>>>>>>
>>>>>>>>> Can you build a cell about 15 x 15 x 15 A or larger to allow you
>>>>>>>>> to extend the range of the hybrid out to 6 or 7 A?
>>>>>>>>>
>>>>>>>>> You are truncating at about 5.25 A, which might not be enough to
>>>>>>>>> be fully converged (note VASP won't be doing this as it works in K space).
>>>>>>>>>
>>>>>>>>> Matt
>>>>>>>>>
>>>>>>>>> On Thursday, September 21, 2017 at 8:36:29 AM UTC+8, Xiaoming Wang
>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>> Hi Ling,
>>>>>>>>>>
>>>>>>>>>> Thanks for your comments. I have attached my benchmark results
>>>>>>>>>> here.  Based on my tests, it seems that the ADMM basis set size is not so
>>>>>>>>>> important for ionic crystals. One can get reasonable results even with
>>>>>>>>>> smallest basis sets. Btw, the cutoff and rel_cutoff used are well converged
>>>>>>>>>> values for my target property. So I think there is still room to improve
>>>>>>>>>> the ADMM basis for Ag and In (with semicore d states in the valence). To
>>>>>>>>>> check whether it is the problem of Ag and In ADMM basis, the simplest way
>>>>>>>>>> is to do the PBE0 calculations without ADMM. But the calculation takes me
>>>>>>>>>> too long time, as also pointed out by Juerg. I suspect if it is possible to
>>>>>>>>>> do HFX without ADMM using MOLOPT basis sets.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Best,
>>>>>>>>>> Xiaoming
>>>>>>>>>>
>>>>>>>>>> On Wednesday, September 20, 2017 at 6:20:09 AM UTC-4, S Ling
>>>>>>>>>> wrote:
>>>>>>>>>>>
>>>>>>>>>>> Hi
>>>>>>>>>>>
>>>>>>>>>>> I am not sure whether you are using the same ADMM basis sets
>>>>>>>>>>> (see your initial input) for these new test calculations. One thing which I
>>>>>>>>>>> can see is that you are not using the largest available ADMM basis sets for
>>>>>>>>>>> some of the elements, e.g. the largest available ADMM basis sets for Cl, Ag
>>>>>>>>>>> and In are pFIT3, FIT12 and FIT13, respectively. Taking In as an example,
>>>>>>>>>>> the FIT13 ADMM basis set of In contains more p and d functions than FIT11,
>>>>>>>>>>> which may be important for your system.
>>>>>>>>>>>
>>>>>>>>>>> You mentioned a few different functionals, including PBE, PBE0
>>>>>>>>>>> and HSE06, and you have run quite a lot of benchmark tests. It would help
>>>>>>>>>>> if you can tabulate all the numbers you have got (including the reference),
>>>>>>>>>>> so we can understand your problem better.
>>>>>>>>>>>
>>>>>>>>>>> In addition, I can see you're using a CUTOFF of 250 Ry. Please
>>>>>>>>>>> also check whether your calculation is converged with respect to this
>>>>>>>>>>> parameter.
>>>>>>>>>>>
>>>>>>>>>>> Please also keep in mind that CP2K and VASP use different
>>>>>>>>>>> pseudopotentials and basis sets. I wouldn't expect the two codes to give
>>>>>>>>>>> the same numbers for your target properties. If you look into literatures,
>>>>>>>>>>> you will also find people reporting different numbers for the same property
>>>>>>>>>>> using the same method and code.
>>>>>>>>>>>
>>>>>>>>>>> SL
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> --
>>>>>> You received this message because you are subscribed to the Google
>>>>>> Groups "cp2k" group.
>>>>>> To unsubscribe from this group and stop receiving emails from it,
>>>>>> send an email to cp2k+... at googlegroups.com.
>>>>>> To post to this group, send email to cp... at googlegroups.com.
>>>>>> Visit this group at https://groups.google.com/group/cp2k.
>>>>>> For more options, visit https://groups.google.com/d/optout.
>>>>>>
>>>>>
>>>>> --
>>> You received this message because you are subscribed to the Google
>>> Groups "cp2k" group.
>>> To unsubscribe from this group and stop receiving emails from it, send
>>> an email to cp2k+... at googlegroups.com.
>>> To post to this group, send email to cp... at googlegroups.com.
>>> Visit this group at https://groups.google.com/group/cp2k.
>>> For more options, visit https://groups.google.com/d/optout.
>>>
>>
>> --
> You received this message because you are subscribed to the Google Groups
> "cp2k" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to cp2k+uns... at googlegroups.com.
> To post to this group, send email to cp... at googlegroups.com.
> Visit this group at https://groups.google.com/group/cp2k.
> For more options, visit https://groups.google.com/d/optout.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20170921/dfe9d28f/attachment.htm>