[CP2K-user] [CP2K:17627] Re: Large discrepancy in xTB results from CP2K vs DFTB+
Xavier Bidault
jazzquark at gmail.com
Wed Sep 7 16:05:01 UTC 2022
So I won't consider this xTB issue fixed. Yet, I have no issue with
EPS_DEFAULT when I use DFTB2 or DFTB3(diag) in CP2K for variable-cell
optimization. Something is still off with xTB. But I must say I lack time
to really look into it.
I also refer to this post (
https://groups.google.com/g/cp2k/c/lESdjFfgq3g/m/drfONg10AQAJ). Actually I
won't be looking for a straight comparison between CP2K and DFTB+, but
rather for a steady/stable/reliable behavior of xTB at CP2K first (DFTB+ has
other issues about the optimization algorithm, IMO). If of any help, please
find attached my input script for beta-HMX using xTB at CP2K. Poisson/Ewald
parameters are not enforced, but does CP2K set them up differently if
EPS_DEFAULT is different? And why different EPS_DEFAULT would result in
such large volume variations and unconverged behavior (see attached figure,
not only for beta-HMX, and using supercells to have dimensions > 10
angstroms). The energy seems to converge for very low EPS_DEFAULT, but some
spurious angle/volume variations remain. Similarly, I can show that xTB is
not reliable for Molecular Dynamics, which prevents relevant prediction of
the behavior under high pressure/temperature.
[image: DFTB_20220722_slide6.jpg]
On Wed, Sep 7, 2022 at 9:42 AM Magnus Rahm <magnus at compulartech.com> wrote:
> If you downloaded or cloned the repo at d56aa67 the bug fix is included.
> Safest approach to compare the versions is probably not a cell optimization
> but a single-point calculation (RUN_TYPE ENERGY_FORCE). The EV-curves in my
> post indicate that cell relaxation with the new version would (probably
> erroneously) have led to a very small cell as well, even though the
> underlying energies have changed significantly.
>
> Kind regards,
> Magnus Rahm
> On Wednesday, September 7, 2022 at 3:11:08 PM UTC+2 jazz... at gmail.com
> wrote:
>
>> >> Could you remind me how to update CP2K 2022.1 to include this bug fix?
>>
>> > I think the easiest approach is to use Docker (following these
>> instructions: https://github.com/cp2k/cp2k/tree/master/tools/docker),
>> unless you want to clone the CP2K repo from github and compile from scratch.
>>
>> I tried yesterday with git clone and started from scratch. The downloaded
>> version (git:d56aa67) should have the xTB bug fix mentioned above (?), but
>> when I run variable-cell optimizations of beta-HMX with various EPS_DEFAULT
>> I find the exact same results as with CP2K 7.1. What's the right git
>> version for this bug fix? Should I copy the git folder instead of using git
>> clone? I mean, how to get the Trunk version you're talking about?
>>
>> On Wed, Sep 7, 2022 at 7:20 AM Magnus Rahm <mag... at compulartech.com>
>> wrote:
>>
>>> Btw, I can confirm that the energy now converges with EPS_DEFAULT also
>>> for the LiO2 system, although the convergence is perhaps a bit slower
>>> (=very small EPS_DEFAULT values needed) than what one might have expected
>>> (see LiO2-EPS_DEFAULT.pdf). The figure I attached in my previous post was
>>> made with EPS_DEFAULT at the default value, if I use 1e-24 I get a bit
>>> closer to DFTB+ but still there is a weird slope in the E-V curve
>>> (EV-LiO2.pdf).
>>>
>>> Furthermore, I had a look at the energy broken down into its different
>>> contributions as a function of volume (LiO2-energies-split.pdf), and FWIW
>>> it indicates that the electronic energy is responsible for the unexpected
>>> slope in the E-V curve (perhaps that was already obvious?).
>>>
>>> > Could you remind me how to update CP2K 2022.1 to include this bug fix?
>>>
>>> I think the easiest approach is to use Docker (following these
>>> instructions: https://github.com/cp2k/cp2k/tree/master/tools/docker),
>>> unless you want to clone the CP2K repo from github and compile from scratch.
>>>
>>> Kind regards,
>>> Magnus Rahm
>>>
>>>
>>> On Wednesday, September 7, 2022 at 2:16:23 AM UTC+2 jazz... at gmail.com
>>> wrote:
>>>
>>>> Thank you. Could you remind me how to update CP2K 2022.1 to include
>>>> this bug fix?
>>>>
>>>> On Tue, Sep 6, 2022 at 10:36 AM Jürg Hutter <hut... at chem.uzh.ch> wrote:
>>>>
>>>>> Hi
>>>>>
>>>>> the updates are now on Github (Trunk version).
>>>>> This should at least fix the strange behavior for changes of
>>>>> EPS_DEFAULT.
>>>>>
>>>>> regards
>>>>>
>>>>> JH
>>>>>
>>>>> ________________________________________
>>>>> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of
>>>>> Jürg Hutter <hut... at chem.uzh.ch>
>>>>> Sent: Tuesday, September 6, 2022 11:37 AM
>>>>> To: cp... at googlegroups.com
>>>>> Subject: Re: [CP2K:17614] Re: Large discrepancy in xTB results from
>>>>> CP2K vs DFTB+
>>>>>
>>>>> Hi
>>>>>
>>>>> I think I found the problem. This is in fact a bug in CP2K and is
>>>>> related to the damping of
>>>>> the "short range" part of the Coulomb term. As mentioned before this
>>>>> short range part
>>>>> is not short range at all, even diverging in periodic systems. We use
>>>>> a damping function
>>>>> for this term and the radius is taken from the range of the basis
>>>>> function on each atom.
>>>>> The bug is now, that this range is not a constant but depends on
>>>>> EPS_DEFAULT.
>>>>> I will work on a solution, but at least the default settings will
>>>>> cause considerable changes
>>>>> in the energies of periodic systems.
>>>>>
>>>>> regards
>>>>>
>>>>> JH
>>>>>
>>>>> ________________________________________
>>>>> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of
>>>>> Magnus Rahm <mag... at compulartech.com>
>>>>> Sent: Monday, September 5, 2022 3:19 PM
>>>>> To: cp2k
>>>>> Subject: Re: [CP2K:17609] Re: Large discrepancy in xTB results from
>>>>> CP2K vs DFTB+
>>>>>
>>>>> Hi,
>>>>>
>>>>> Thank you for valuable input! Here's a breakdown of energies for a
>>>>> periodic LiO2 system (where CP2K and DFTB+ disagree).
>>>>> CP2K:
>>>>>
>>>>> Core Hamiltonian energy:
>>>>> -609.45757320827579
>>>>> Repulsive potential energy:
>>>>> 2.86335541921533
>>>>> Electronic energy:
>>>>> -65.73940900376786
>>>>> DFTB3 3rd order energy:
>>>>> 9.00274299587460
>>>>> Dispersion energy:
>>>>> -2.00065978643714
>>>>> Correction for halogen bonding:
>>>>> 0.00000000000000
>>>>>
>>>>> Total energy:
>>>>> -665.33154358339084
>>>>>
>>>>> outer SCF iter = 1 RMS gradient = 0.49E-06 energy = -665.
>>>>> 3315435834 <(331)%20543-5834>
>>>>> outer SCF loop converged in 1 iterations or 10 steps
>>>>>
>>>>> And the same system with DFTB+ (I don't know this is the best
>>>>> breakdown I can get from DFTB+? This info is from detailed.out.):
>>>>>
>>>>> Fermi level: -0.1574062769 H -4.2832
>>>>> eV
>>>>> Band energy: -254.9890864567 H -6938.6061
>>>>> eV
>>>>> TS: 0.0000000000 H 0.0000
>>>>> eV
>>>>> Band free energy (E-TS): -254.9890864567 H -6938.6061
>>>>> eV
>>>>> Extrapolated E(0K): -254.9890864567 H -6938.6061
>>>>> eV
>>>>> Input / Output electrons (q): 864.0000000000 864.0000000000
>>>>>
>>>>> Energy H0: -610.3586854777 H -16608.7049
>>>>> eV
>>>>> Energy SCC: 13.1915555608 H 358.9605
>>>>> eV
>>>>> Total Electronic energy: -597.1671299169 H -16249.7444
>>>>> eV
>>>>> Repulsive energy: 0.0000000000 H 0.0000
>>>>> eV
>>>>> Total energy: -597.1671299169 H -16249.7444
>>>>> eV
>>>>> Extrapolated to 0: -597.1671299169 H -16249.7444
>>>>> eV
>>>>> Total Mermin free energy: -597.1671299169 H -16249.7444
>>>>> eV
>>>>> Force related energy: -597.1671299169 H -16249.7444
>>>>> eV
>>>>>
>>>>>
>>>>> ----------------------------------------------------------------------------------------------------------------
>>>>> For reference, here are the equivalent breakdowns for the LiF
>>>>> molecule, where the total energies do match quite well.
>>>>> CP2K:
>>>>>
>>>>> Core Hamiltonian energy:
>>>>> -5.57594122418510
>>>>> Repulsive potential energy:
>>>>> 0.00036401843654
>>>>> Electronic energy:
>>>>> 0.08477836575096
>>>>> DFTB3 3rd order energy:
>>>>> -0.00385103760005
>>>>> Dispersion energy:
>>>>> -0.00008325087778
>>>>> Correction for halogen bonding:
>>>>> 0.00000000000000
>>>>>
>>>>> Total energy:
>>>>> -5.49473312847544
>>>>>
>>>>> outer SCF iter = 1 RMS gradient = 0.12E-06 energy =
>>>>> -5.4947331285
>>>>> outer SCF loop converged in 1 iterations or 25 steps
>>>>>
>>>>> DFTB+
>>>>> Fermi level: -0.3434874008 <(343)%20487-4008>
>>>>> H -9.3468 eV
>>>>> Band energy: -3.7493389034 H -102.0247
>>>>> eV
>>>>> TS: 0.0000000000 H 0.0000
>>>>> eV
>>>>> Band free energy (E-TS): -3.7493389034 H -102.0247
>>>>> eV
>>>>> Extrapolated E(0K): -3.7493389034 H -102.0247
>>>>> eV
>>>>> Input / Output electrons (q): 8.0000000444 8.0000000000
>>>>>
>>>>> Energy H0: -5.5743451431 <(574)%20345-1431>
>>>>> H -151.6856 eV
>>>>> Energy SCC: 0.0807122067 H 2.1963
>>>>> eV
>>>>> Total Electronic energy: -5.4936329365 H -149.4894
>>>>> eV
>>>>> Repulsive energy: 0.0000000000 H 0.0000
>>>>> eV
>>>>> Total energy: -5.4936329365 H -149.4894
>>>>> eV
>>>>> Extrapolated to 0: -5.4936329365 H -149.4894
>>>>> eV
>>>>> Total Mermin free energy: -5.4936329365 H -149.4894
>>>>> eV
>>>>> Force related energy: -5.4936329365 H -149.4894
>>>>> eV
>>>>>
>>>>>
>>>>> ----------------------------------------------------------------------------------------------------------------
>>>>>
>>>>> > I recently run variable-cell optimization of various molecular
>>>>> crystals and I found xTB at CP2K ultra sensitive to EPS_DEFAULT. Tested
>>>>> from 1e-5 to 1e-24 (with EPS_SCF 1e-8), and no convergence happened.
>>>>>
>>>>> Thank you for sharing this info! I tried a series of calculations with
>>>>> LiO2 using varying values of EPS_DEFAULT (using default EPS_SCF) and found
>>>>> the same effect; no convergence with EPS_DEFAULT (or perhaps unreasonably
>>>>> slow convergence). I attach a figure showing these results, including the
>>>>> energy broken down into the different parts as specified in the CP2K
>>>>> output. Note the energy scale, the changes with EPS_DEFAULT are really
>>>>> quite substantial. In the LiF (non-PBC) case, the corresponding curves look
>>>>> completely flat on the same scale. I don't know what to make of this
>>>>> result, but perhaps someone else does?
>>>>>
>>>>> Magnus
>>>>>
>>>>> [X]
>>>>> On Monday, September 5, 2022 at 12:18:26 PM UTC+2 jazz... at gmail.com
>>>>> wrote:
>>>>> I recently run variable-cell optimization of various molecular
>>>>> crystals and I found xTB at CP2K ultra sensitive to EPS_DEFAULT. Tested
>>>>> from 1e-5 to 1e-24 (with EPS_SCF 1e-8), and no convergence happened. I just
>>>>> ended up with EPS_DEFAULT 1e-10 as a "gut" choice. Also, the behavior of
>>>>> xTB at CP2K is doutfull with MD even at ambiant conditions, where the
>>>>> converged volume is barely larget than at 0K. Depending on EPS_DEFAULT, it
>>>>> can even be smaller at ambient T. Weird. The behavior of DFTB2 at CP2K
>>>>> is far better.
>>>>>
>>>>> I found that DFTB+ has other issues. xTB at DFTB+ has no convergence
>>>>> issue, but the recommended variable-cell optimization algorithm has flaws.
>>>>> The unit cell and a supercell does NOT always end up with related lattice
>>>>> parameters. The main issue is that some 90° angles are not preserved with
>>>>> DFTB+ whereas CP2K does (with no symmetry enforced, obviously). Some
>>>>> inconsistencies appears in DFTB+ with a lattice dimensions < 10 angstroms
>>>>> in the unit cell versus > 10 angstroms in the supercell. A proper tight
>>>>> mesh of k-points does not improve. So I'm afraid that xTB at DFTB+ (or
>>>>> DFTB+, actually) cannot be a relevant choice for crystal structure
>>>>> predictions, for instance.
>>>>>
>>>>> xTB may be unreliable with CP2K and DFTB+, but for the different
>>>>> reasons above. You can check these weird behaviors with your own crystals
>>>>> of interest.
>>>>>
>>>>> Xavier
>>>>>
>>>>>
>>>>>
>>>>> Le lun. 5 sept. 2022, 3:59 AM, Jürg Hutter <hut... at chem.uzh.ch> a
>>>>> écrit :
>>>>> Hi
>>>>>
>>>>> thank you for testing. Could you send a break down of the energies for
>>>>> the LiF molecule for
>>>>> the two codes? That might help to recognize the source of the
>>>>> difference.
>>>>>
>>>>> regards
>>>>>
>>>>> JH
>>>>>
>>>>> ________________________________________
>>>>> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of
>>>>> Magnus Rahm <mag... at compulartech.com>
>>>>> Sent: Monday, September 5, 2022 8:40 AM
>>>>> To: cp2k
>>>>> Subject: [CP2K:17599] Re: Large discrepancy in xTB results from CP2K
>>>>> vs DFTB+
>>>>>
>>>>> For the record, the problem is the same in CP2K version 2022.1.
>>>>>
>>>>> On Thursday, September 1, 2022 at 12:48:35 PM UTC+2 Magnus Rahm wrote:
>>>>> Dear all,
>>>>>
>>>>> I want to use CP2K (version 8.2, trying to get a more recent version
>>>>> compiled) together with xTB for a crystal containing Li and O. I get
>>>>> strange results already for a simple LiO2 crystal:
>>>>>
>>>>> * There is a very large discrepancy compared to DFTB+ (version 22.1).
>>>>> * Mulliken charges tend to be large, meaning that CHECK_ATOMIC_CHARGES
>>>>> stops the SCF. If I turn it off, the system tends to converge
>>>>> systematically to values just outside the "chemical range". Mulliken
>>>>> charges obtained by DFTB+ are significantly smaller (and within "chemical
>>>>> range").
>>>>> * The energy-volume curve looks strange and very different from DFTB+.
>>>>>
>>>>> I have tried converging with respect to system size and the EWALD /
>>>>> ALPHA and GMAX parameters, but they have only a marginal impact. I have
>>>>> tried similar calculations for a number of periodic systems. Sometimes I
>>>>> get agreement, sometimes not. I also tried calculations for CO and NO
>>>>> molecules which agree perfectly between CP2K and DFTB+, whereas an
>>>>> artificial LiF molecule does not.
>>>>>
>>>>> A perhaps related issue was reported in
>>>>> https://groups.google.com/g/cp2k/c/oFwgGcQuySs but the solutions
>>>>> suggested there did not solve my problem.
>>>>>
>>>>> I attach input scripts for CP2K and DFTB+, as well as a figure showing
>>>>> the E-V curve for LiO2 obtained with CP2K and DFTB+. I'm new to CP2K, DFTB+
>>>>> and xTB so I suspect I have made some simple mistake, and any advice is
>>>>> appreciated.
>>>>>
>>>>> Kind regards,
>>>>> Magnus Rahm
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> 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<mailto:
>>>>> cp2k+uns... at googlegroups.com>.
>>>>> To view this discussion on the web visit
>>>>> https://groups.google.com/d/msgid/cp2k/c51e70a2-7f7a-4887-8ada-971d840a1b13n%40googlegroups.com
>>>>> <
>>>>> https://groups.google.com/d/msgid/cp2k/c51e70a2-7f7a-4887-8ada-971d840a1b13n%40googlegroups.com?utm_medium=email&utm_source=footer
>>>>> >.
>>>>>
>>>>> --
>>>>> 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 view this discussion on the web visit
>>>>> https://groups.google.com/d/msgid/cp2k/ZR0P278MB075983A13AA78F53FFB422559F7F9%40ZR0P278MB0759.CHEP278.PROD.OUTLOOK.COM
>>>>> .
>>>>>
>>>>> --
>>>>> 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<mailto:
>>>>> cp2k+uns... at googlegroups.com>.
>>>>> To view this discussion on the web visit
>>>>> https://groups.google.com/d/msgid/cp2k/7d455fce-a22e-44f1-ac48-18495f9553a5n%40googlegroups.com
>>>>> <
>>>>> https://groups.google.com/d/msgid/cp2k/7d455fce-a22e-44f1-ac48-18495f9553a5n%40googlegroups.com?utm_medium=email&utm_source=footer
>>>>> >.
>>>>>
>>>>> --
>>>>> 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 view this discussion on the web visit
>>>>> https://groups.google.com/d/msgid/cp2k/ZR0P278MB0759EEC38142F9D207742F549F7E9%40ZR0P278MB0759.CHEP278.PROD.OUTLOOK.COM
>>>>> .
>>>>>
>>>>> --
>>>>> 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 view this discussion on the web visit
>>>>> https://groups.google.com/d/msgid/cp2k/ZR0P278MB0759A3BAE12A4F8C0298072A9F7E9%40ZR0P278MB0759.CHEP278.PROD.OUTLOOK.COM
>>>>> .
>>>>>
>>>> --
>>> 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 view this discussion on the web visit
>>> https://groups.google.com/d/msgid/cp2k/1234148a-7a9a-4c6b-a0c9-13caabfe8811n%40googlegroups.com
>>> <https://groups.google.com/d/msgid/cp2k/1234148a-7a9a-4c6b-a0c9-13caabfe8811n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>> .
>>>
>> --
> 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+unsubscribe at googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/cp2k/e80f7efb-e11f-4e6c-bc4a-a20e0e234b57n%40googlegroups.com
> <https://groups.google.com/d/msgid/cp2k/e80f7efb-e11f-4e6c-bc4a-a20e0e234b57n%40googlegroups.com?utm_medium=email&utm_source=footer>
> .
>
--
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+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/CAD0N%2BNV1dZSDgzWgRSpzqq8PEX_EqMK1bMP7_P8BUn0EkrTTEA%40mail.gmail.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220907/9069e11a/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: DFTB_20220722_slide6.jpg
Type: image/jpeg
Size: 284919 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220907/9069e11a/attachment-0001.jpg>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 2_vc-relax.in
Type: application/octet-stream
Size: 6486 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220907/9069e11a/attachment-0001.obj>
More information about the CP2K-user
mailing list