[CP2K-user] [CP2K:22444] Question about Al (GTH-SCAN-q3) forces with r2SCAN compared with all-electron / PAW results
liu wei
whisper256136 at gmail.com
Wed Oct 14 05:39:16 UTC 2026
Dear Prof. Thomas Kühne,
Thank you for repeating the QZVPP calculations and for the suggestions. I
have now run the checks you proposed, all
on the same initial structure (302 atoms, 2 Al) with the same settings as
before (GAPW + GAPW_ACCURATE_XCINT,
800 Ry / REL_CUTOFF 60). Only the Al kind is changed between runs; O, P and
Zn always use GTH-SCAN with
TZV2P-MOLOPT-SCAN. The all-electron Al reference is QZVPP-MOLOPT-PBE-ae
with POTENTIAL ALL. Below, Al |ΔF| is the
RMS over the two Al atoms of the vector force difference, in meV/Å.
- Finite-difference check
RUN_TYPE DEBUG (forces only, DX 0.001 bohr, EPS_SCF 1e-7) for the two Al
components with the largest GTH vs
all-electron difference (atom 62 z: 212 meV/Å, atom 70 x: −235 meV/Å).
Forces in meV/Å:
| Al | Component | Numerical | Analytical | Difference |
|---|---|---|---|---|
| GTH-SCAN-q3 | 62 z | −1153.20 | −1153.21 | −0.01 |
| GTH-SCAN-q3 | 70 x | −21.46 | −21.68 | −0.22 |
| all-electron | 62 z | −1365.27 | −1365.22 | 0.05 |
| all-electron | 70 x | 213.73 | 213.44 | −0.29 |
The analytical forces agree with the energies to within 0.3 meV/Å on both
sides, so the difference comes from the two
energy surfaces themselves, not from the force evaluation.
- GTH basis, GTH-PBE-q3 vs GTH-SCAN-q3, and PBE / r2SCAN / SCAN
For the GTH basis I used the MOLOPT-SR series SZV / DZVP (BASIS_MOLOPT) and
TZVP / TZV2P (BASIS_MOLOPT_UCL), which is
the largest SR series I found. Each basis was combined with GTH-PBE-q3 and
with GTH-SCAN-q3, and each combination was
run with PBE, r2SCAN and SCAN. Each result is compared with the
all-electron Al calculation using the same functional
(D3(BJ) with the matching reference functional). Al |ΔF| relative to
all-electron Al (meV/Å):
| Functional | Al basis | GTH-PBE-q3 | GTH-SCAN-q3 |
|---|---|---|---|
| PBE | SZV-SR | 110 | 204 |
| PBE | DZVP-SR | 103 | 208 |
| PBE | TZVP-SR | 132 | 232 |
| PBE | TZV2P-SR | 149 | 250 |
| r2SCAN | SZV-SR | 127 | 227 |
| r2SCAN | DZVP-SR | 123 | 228 |
| r2SCAN | TZVP-SR | 151 | 251 |
| r2SCAN | TZV2P-SR | 169 | 270 |
| SCAN | SZV-SR | 141 | 214 |
| SCAN | DZVP-SR | 112 | 203 |
| SCAN | TZVP-SR | 139 | 228 |
| SCAN | TZV2P-SR | 157 | 248 |
| r2SCAN | TZV2P-MOLOPT-SCAN | | 259 |
| SCAN | TZV2P-MOLOPT-SCAN | | 237 |
What I read from this:
- With SCAN, which matches the potential, GTH-SCAN-q3 is still 200–250
meV/Å from all-electron, only 6–11% less
than with r2SCAN. So using the SCAN potential with r2SCAN does not
explain the difference.
- GTH-PBE-q3 is closer to all-electron than GTH-SCAN-q3 by 102–117 meV/Å
for every basis and every functional,
including SCAN. This difference hardly depends on the functional, so it
seems to come from the parameters of the
two potentials themselves.
- Even with the fully matched combination (PBE functional with GTH-PBE-q3),
103–149 meV/Å remain. I have not
separated whether this comes from the frozen 2s2p shell or from the
missing nonlinear core correction.
- Enlarging the GTH basis moves the forces away from the all-electron
result in all six series. DZVP or SZV is closest,
TZV2P is farthest, and TZVP → TZV2P still adds 17–20 meV/Å. I read the
better agreement of the small bases as error
cancellation.
These results support your suggestion that the transferability of the Al
potential is the main problem. The limits
are that this is a single structure with two Al atoms, and that O, P and Zn
keep their SCAN potentials in the PBE runs
(the all-electron reference uses the same O, P and Zn, so this should
cancel in the Al comparison).
- Files
The attached cp2k_GTH-SCAN_report_v3.zip contains the input and output of
every run above, with the user name in the
paths replaced by XXXXXXXXXX:
- A_Al/initial/F1_DEBUG_GTH, F2_DEBUG_AE: finite-difference checks
- A_Al/initial/A4_QZVPPae, S0_PBE_AE, S2_SCAN_AE: all-electron Al reference
with r2SCAN / PBE / SCAN
- A_Al/initial/S1_SCAN_GTH: GTH-SCAN-q3 with TZV2P-MOLOPT-SCAN and SCAN
- A_Al/initial/M_<functional>_<basis>_<PBEpp|SCANpp>: the basis / potential
/ functional matrix; the r2SCAN cells
for SZV / TZVP with GTH-SCAN-q3 are B1_SZVSR_SCANpp and B2_TZVPSR_SCANpp.
The r2SCAN DZVP cells and TZV2P with
GTH-SCAN-q3 are A1–A3 in the v2 package.
Thank you again for your time and help.
Best regards,
Wei LIU
在2026年10月14日星期三 UTC+9 04:59:36<Thomas Kühne> 写道：
> Dear Wei,
>
> many thanks for the update. QZVPP-MOLOPT-PBE-ae is a reasonable choice for
> this control calculation. Its optimization
> for PBE does not prevent its use with r2SCAN, although convergence should
> be checked for the property of interest.
>
> I have now completed the corresponding calculations independently and
> reproduce your QZVPP results. In addition,
> changing from TZVPP to QZVPP changes the Al forces by about 13 meV/Å,
> while refining the atom-centered integration
> grid changes the QZVPP forces by only 0.9 meV/Å. Thus, the AE basis
> contributed to the original discrepancy, but does
> not explain most of the remaining difference.
>
> Al pseudopotential transferability is now my main suspicion. Before
> considering a new fit, I would check one or two analytical
> Al force components against finite differences of the total energy and
> systematically enlarge the GTH basis as well.
>
> An additional matched SCAN comparison between GTH and all-electron Al
> would help distinguish transferability of the original
> potential from effects of using it with r2SCAN.
>
> I do not have a validated SCAN q11 potential and matching basis to
> recommend. The MPC-PBE-q11 potential would require
> a separate assessment rather than being a direct replacement.
>
> If you sent me your new input and output files, I can have a look at them
> too ...
>
> Best,
> Thomas
>
> Am 13.10.2026 um 12:40 schrieb liu wei <whispe... at gmail.com>:
>
> Dear Prof. Thomas Kühne,
>
> Thank you very much for checking the UZH data and for running the control
> calculations yourself. The point that Al
> was still treated as GPW in my GTH runs was new to me, and it is
> reassuring that FORCE_PAW / GAPW_XC give the same
> ~306 meV/Å.
>
> **Larger all-electron Al basis**
>
> Following your suggestion, I repeated the all-electron Al calculation on
> the initial structure with a larger basis.
> Everything else was kept the same (GAPW + GAPW_ACCURATE_XCINT, POTENTIAL
> ALL); only the Al basis was changed from
> pob-TZVP-rev2 to QZVPP-MOLOPT-PBE-ae (BASIS_MOLOPT_UZH). Al |ΔF| RMS
> (meV/Å):
>
> | Comparison | pob-TZVP-rev2 | QZVPP-MOLOPT-PBE-ae |
> |---|---|---|
> | GTH-SCAN-q3 vs all-electron Al | 306 | 259 |
> | all-electron Al vs VASP PAW (Al_sv_GW) | 62 | 68 |
>
> - The larger basis changes the Al forces by about 60 meV/Å (other elements
> ≤ 6 meV/Å, pressure 0.08 kbar), so the
> pob-TZVP-rev2 reference was indeed not fully converged, as you expected.
> - However, a difference of about 260 meV/Å between GTH-SCAN-q3 and the
> all-electron Al remains.
>
> I am not sure whether QZVPP-MOLOPT-PBE-ae is an appropriate choice here,
> since it was optimized for PBE and I use it
> with r2SCAN, and I have no larger basis to check whether it is itself
> converged. Would you consider it suitable for
> this test, or would you recommend another all-electron basis?
>
> **Semicore Al potential**
>
> I could not find a semicore (q11) GTH potential for Al with SCAN in the
> CP2K 2026.2 data directory. The only one I
> found is GTH-MPC-PBE-q11 in MPC_POTENTIALS, which is a PBE potential and
> has no matching basis set, so I did not run
> this comparison. If you know of a suitable semicore potential and basis, I
> would be happy to try it.
>
> **Zn**
>
> Thank you for the clarification. I will keep USE_FINER_GRID for
> production, and use GAPW + GAPW_ACCURATE_XCINT +
> USE_FINER_GRID when lower Zn force noise is needed.
>
> I can send the new input and output files if they are useful.
>
> Thank you again for your time and help.
>
> Best regards,
> Wei LIU
>
> 在2026年10月13日星期二 UTC+9 17:10:31<Thomas Kühne> 写道：
>
>> Dear Wei,
>>
>> GAPW with accurate XC integration and USE_FINER_GRID seems to resolve the
>> Zn grid problem,
>> but so far not the Al discrepancy. A smoother Zn potential or basis might
>> reduce the integration cost,
>> but your tests do not establish that it would remove the grid sensitivity.
>>
>> I found no evidence of corruption in the UZH data integration. The
>> relevant entries are numerically
>> unchanged, and the Al potential matches the original SCAN file.
>>
>> One important detail in your output is that Al is still treated as GPW
>> because its basis is entirely soft
>> under the chosen decomposition settings. O and Zn receive the
>> atom-centered GAPW treatment.
>> FORCE_PAW T enables the one-center treatment but does not itself
>> guarantee a nonzero hard-soft
>> density difference. The split depends on EPSFIT and HARD_EXP_RADIUS.
>>
>> I have completed control calculations for the initial structure with the
>> same geometry, basis sets and
>> pseudopotentials, with GAPW_ACCURATE_XCINT T enabled in all runs.
>> Relative to your GAPW
>> baseline, adding FORCE_PAW T or using GAPW_XC with FORCE_PAW T changes
>> the Al forces by
>> less than 1 meV/Å. The difference from your all-electron result remains
>> approximately 306 meV/Å.
>>
>> Finally, Al GTH-SCAN-q3 contains only three explicit valence
>> electrons. Using it with GAPW does not
>> restore the frozen core! An appropriate semicore potential could provide
>> another diagnostic comparison.
>> Checking the basis-set convergence of your existing all-electron Al
>> calculation would also be useful.
>> Checking the basis-set convergence of your all-electron Al reference
>> would also be useful. In our own
>> tests, QZVPP-quality basis sets have sometimes been necessary for
>> reliable all-electron results.
>> Comparing the present pob-TZVP-rev2 basis with a suitable QZVPP-quality
>> basis would therefore be
>> worthwhile. However, Al potential transferability remains a plausible
>> explanation.
>>
>> Greetings,
>> Thomas
>>
>> Am 13.10.2026 um 04:33 schrieb liu wei <whispe... at gmail.com>:
>>
>> Dear Prof. Jürg Hutter,
>>
>> Thank you very much for your quick and helpful reply. I ran the tests you
>> suggested; the inputs, outputs and an
>> updated analysis script are in the attached cp2k_GTH-SCAN_report_v2.zip
>> (new section A3 and an extended B1 in the README).
>> Please correct me if I have misread anything.
>>
>> **1. Al: MOLOPT-SR basis**
>>
>> I kept everything else fixed (GAPW + GAPW_ACCURATE_XCINT, all other
>> elements TZV2P-MOLOPT-SCAN / GTH-SCAN) and changed
>> only the Al kind, on the original structure and the three MD frames. Al
>> |ΔF| RMS against the all-electron Al
>> (meV/Å):
>>
>> | Al basis + potential | initial | Al4 | Al5 | Al6 |
>> |---|---|---|---|---|
>> | TZV2P-MOLOPT-SCAN-GTH-q3 + GTH-SCAN-q3 (before) | 306 | 466 | 351 | 333
>> |
>> | DZVP-MOLOPT-SR-GTH-q3 + GTH-SCAN-q3 | 275 | 410 | 307 | 281 |
>> | TZV2P-MOLOPT-SR-GTH-q3 (BASIS_MOLOPT_UCL) + GTH-SCAN-q3 | 314 | 487 |
>> 368 | 350 |
>> | DZVP-MOLOPT-SR-GTH-q3 + GTH-PBE-q3 | 172 | 245 | 187 | 181 |
>>
>> - With the SR basis the Al difference changes only a little (about −10 %
>> with DZVP-SR, no change with TZV2P-SR), so it
>> does not seem to come from the SCAN basis.
>> - Surprisingly to me, switching only the potential from GTH-SCAN-q3 to
>> GTH-PBE-q3 removes about half of the difference,
>> in the direction of the all-electron result, although the two parameter
>> sets differ by less than 1 %. The trend on
>> the Al–O pairs and on the relative energies of the Al4/5/6 frames is
>> reduced in the same way, but not removed
>> (about 180 meV/Å remains).
>>
>> So the Al forces seem to be very sensitive to the potential parameters.
>> Would you expect this for GTH-SCAN-q3, and
>> would you recommend a different approach for Al in oxides with r2SCAN,
>> for example a potential with 2s2p in the
>> valence (I could not find one for SCAN in POTENTIAL_UZH), or simply the
>> all-electron Al with GAPW?
>>
>> **2/3. Zn grid sensitivity**
>>
>> Following your suggestion, I added GAPW, GAPW_ACCURATE_XCINT and
>> USE_FINER_GRID step by step (1100 Ry, 302 atoms,
>> 4 rigid shifts). Zn force spread, |ΔF| RMS (meV/Å):
>>
>> | Setting | Zn spread |
>> |---|---|
>> | GPW | ~390 |
>> | GAPW + GAPW_ACCURATE_XCINT | 37 |
>> | GPW + USE_FINER_GRID | 28–32 |
>> | GAPW + GAPW_ACCURATE_XCINT + USE_FINER_GRID | 0.7–0.8 |
>>
>> - The full combination works very well: the Zn noise drops below 1 meV/Å.
>> Thank you for this.
>> - I also changed only the Zn basis to TZVP-MOLOPT-SR-GTH and the
>> potential to GTH-PBE-q12; the noise stayed at
>> ~390 meV/Å, so with these choices it seems to come from r2SCAN on the
>> grid rather than from the SCAN basis or potential.
>>
>> May I ask one more question on this? Would it be possible, or sensible at
>> all, to reduce this grid sensitivity on the
>> side of the potential or the basis set, for example a Zn potential/basis
>> built to be smoother on the grid for
>> meta-GGAs? Since the SR basis and the GTH-PBE potential did not change
>> it, I do not know whether this is something the
>> potential or basis can address.
>>
>> Thank you again for your time and help.
>>
>> Best regards,
>> Wei LIU
>>
>>
>> 在2026年10月12日星期一 UTC+9 22:23:26<Jürg Hutter> 写道：
>>
>>> Hi
>>>
>>> 1) You can try the Al (sr) basis from BASIS_MOLOPT. This text would give
>>> a hint if the UZH SCAN basis for Al is in some
>>> way corrupted. The SCAN/Hybrid basis sets are less texted than the PBE
>>> ones.
>>> 2/3) r2SCAN needs higher cutoffs than PBE. If the force wiggles are too
>>> large you can try GAPW and accurate_intrgration and
>>> finer_grid (in that order, accumulative)
>>>
>>> regards
>>> JH
>>>
>>> ________________________________________
>>> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of liu
>>> wei <whispe... at gmail.com>
>>> Sent: Monday, October 12, 2026 2:15 PM
>>> To: cp2k
>>> Subject: [CP2K:22433] Question about Al (GTH-SCAN-q3) forces with r2SCAN
>>> compared with all-electron / PAW results
>>>
>>> Dear CP2K developers and users,
>>>
>>> I am a student working on ab initio MD of a ZnO–P2O5–Al2O3 glass (302
>>> atoms, Γ point) with r2SCAN-D3(BJ),
>>> the GTH-SCAN potentials and the MOLOPT-SCAN basis sets (CP2K 2026.2).
>>> While I was checking my forces against VASP,
>>> I noticed a difference on Al that I do not understand, and I would be
>>> very grateful for your advice. I may well be
>>> missing something in my setup, so please correct me if so.
>>>
>>> I have attached all inputs and outputs, a short analysis script (numpy
>>> only) and a README with the full tables
>>> (cp2k_GTH-SCAN_report.zip).
>>>
>>> **1. Al forces**
>>>
>>> On one 302-atom structure I obtained the following force differences
>>> (|ΔF| RMS per element, meV/Å):
>>>
>>> | A − B | Zn | O | P | Al |
>>> |---|---|---|---|---|
>>> | CP2K (GPW, TZV2P-MOLOPT-SCAN, GTH-SCAN, 800 Ry, FINER) − VASP (PAW
>>> O_h/P_h/Zn/Al_sv_GW, 1300 eV) | 41 | 55 | 61 | 320 |
>>> | CP2K GAPW, Al → ALL + pob-TZVP-rev2 − CP2K GAPW, Al GTH-SCAN-q3
>>> (everything else identical) | 3 | 40 | 11 | 306 |
>>> | CP2K GAPW with all-electron Al − VASP Al_sv_GW | 58 | 39 | 61 | 62 |
>>>
>>> - For Zn, O and P the two codes agree to about 40–60 meV/Å, and the
>>> pressure agrees to within 1 kbar, but for Al the
>>> difference is about 300 meV/Å.
>>> - When I replaced only the Al GTH-SCAN-q3 by an all-electron Al in GAPW,
>>> most of this difference (about 94 %)
>>> disappeared in my test.
>>> - I tried to check a few other possible reasons: USE_FINER_GRID, GPW vs
>>> GAPW (0.6 meV/Å
>>> on Al), and the frozen 2s2p core on the VASP side (Al vs Al_sv_GW
>>> changes the Al forces by 69 meV/Å, in the other
>>> direction). None of them seemed to explain it, but I am not sure whether
>>> I checked them in the right way.
>>> - I also looked at three MD snapshots with 4-, 5- and 6-coordinated Al.
>>> On all 6 Al the difference seems to follow the
>>> same trend: compared with the all-electron Al, GTH-SCAN-q3 appears to
>>> pull the short Al–O bonds (< 1.85 Å) in
>>> and push the longer Al–O contacts (2.1 Å and beyond) out; between 1.85
>>> and 2.1 Å the sign varies. Here I project
>>> ΔF = F(all-electron) − F(GTH-SCAN-q3) on each Al–O pair as (ΔF_O −
>>> ΔF_Al)·u, with u the unit vector from Al to O.
>>> A positive value means the all-electron force along the pair is more
>>> repulsive than the GTH-SCAN-q3 one (so
>>> GTH-SCAN-q3 pulls the pair together), and a negative value means it is
>>> more attractive (GTH-SCAN-q3 pushes the
>>> pair apart). These are differences, not the total forces. For example, I
>>> get +812 meV/Å for a bond of 1.66 Å and
>>> −436 meV/Å for a contact of 2.10 Å.
>>> - One limitation I am aware of: the all-electron run also uses a
>>> different Al basis (pob-TZVP-rev2), so I could not
>>> separate the effect of the potential from that of the basis. Changing
>>> only the MOLOPT basis (TZVP → TZV2P) moved
>>> the Al forces by 87 meV/Å, so I guessed the potential plays the larger
>>> role, but I am not confident about this.
>>>
>>> **2. Zn (a smaller question)**
>>>
>>> With GPW, r2SCAN + TZVP-MOLOPT-SCAN + GTH-SCAN-q12 gave a Zn force
>>> spread of about 390 meV/Å RMS over 4 rigid shifts
>>> of the same structure, while PBEsol + TZVP-MOLOPT-SR + GTH-PBE-q12
>>> changed by only about 2 meV/Å when the FFT grid
>>> changed (1100 vs 1350 Ry). USE_FINER_GRID reduced the r2SCAN spread to
>>> about 30 meV/Å, which works well for me.
>>> I did not separate the effects of the functional, the basis and the
>>> potential, so I am not sure where this comes from.
>>>
>>> **My questions**
>>>
>>> 1. Could you help me understand what might cause the Al difference? Is
>>> it something one would expect with
>>> GTH-SCAN-q3 (for example, because it was fitted for SCAN rather than
>>> r2SCAN), or is there likely a problem in my
>>> setup or in the way I compared the results?
>>> 2. For Al in oxides with SCAN/r2SCAN, would you recommend a different
>>> potential or basis (for example, one with
>>> 2s2p in the valence), or another way to check this more carefully?
>>> 3. For Zn with r2SCAN, is this grid sensitivity expected, and is there
>>> anything you would recommend besides
>>> USE_FINER_GRID?
>>>
>>> I would be happy to run any additional tests you suggest. Thank you very
>>> much for your time and for developing
>>> CP2K.
>>>
>>> Best regards,
>>> Wei LIU
>>>
>>>
>>> --
>>> 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...@
>>> googlegroups.com>.
>>> To view this discussion visit
>>> https://groups.google.com/d/msgid/cp2k/189080b0-6d63-4515-9a04-191b65347daan%40googlegroups.com
>>> <
>>> https://groups.google.com/d/msgid/cp2k/189080b0-6d63-4515-9a04-191b65347daan%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 visit
>> https://groups.google.com/d/msgid/cp2k/f6b35e26-6305-4de1-abb2-c135735ab583n%40googlegroups.com
>> <https://groups.google.com/d/msgid/cp2k/f6b35e26-6305-4de1-abb2-c135735ab583n%40googlegroups.com?utm_medium=email&utm_source=footer>
>> .
>> <cp2k_GTH-SCAN_report_v2.zip>
>>
>>
>>
> --
> 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 visit
> https://groups.google.com/d/msgid/cp2k/86b7344d-bced-4a03-8513-2cd6eee347e0n%40googlegroups.com
> <https://groups.google.com/d/msgid/cp2k/86b7344d-bced-4a03-8513-2cd6eee347e0n%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 visit https://groups.google.com/d/msgid/cp2k/12245f53-7b23-4970-b4cb-c5848fcac3b4n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20261013/685f98e5/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cp2k_GTH-SCAN_report_v3.zip
Type: application/x-zip
Size: 1299931 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20261013/685f98e5/attachment-0001.bin>
More information about the CP2K-user
mailing list