[CP2K-user] [CP2K:18045] Thoughts on the instability of HSE06+ADMM
Nicholas Winner
nwinner at berkeley.edu
Wed Nov 16 17:24:47 UTC 2022
Hi Matt. It looks like the reason yours converged and mine didn't is the
auxiliary basis. I was running comparing the admm-{dz, dzp, tzp} basis sets
from the new BASIS_ADMM_UZH file. For that case I was eventually able to
get dz quality converged only if I removed the initial density screening,
while the others diverged like I said. These basis sets, however, seem to
have a worse condition number and more diffuse functions than the FIT type.
Using fit9 and cfit10 like you did, 360Gb of integrals are reduced to 120Gb
of integrals, and the convergence is good again. I'll have to be a little
more careful with which auxiliary basis I choose, I suppose.
HSE always complains about the box being too small. I think its a bug? So
long as the minimum distance between images is 10A it should be fine since
the decay length is 5A.
On Wednesday, November 16, 2022 at 2:29:13 AM UTC-8 mattwa... at gmail.com
wrote:
> Hi,
> seemed to work OK for me running a 3x3x3 of the conventional cell. See
> attached starting from PBE guess.
> Still HSE complains about box being too small ... but convergence is
> smooth.
> Matt
> On Tuesday, 15 November 2022 at 17:44:24 UTC nwi... at berkeley.edu wrote:
>
>> I was not able to get convergence using algo strict. In fact, I always
>> thought that IRAC was a little more flexible for optimization than strict
>> orthogonality. Maybe I'm wrong.
>>
>> CG including 3PNT (where I turned off rotation) also does not work.
>> On Friday, November 11, 2022 at 12:52:55 AM UTC-8 Matthias Krack wrote:
>>
>>> Did you try already OT MINIMIZER CG (and possibly no ALGORITHM IRAC but
>>> the default STRICT)?
>>>
>>>
>>>
>>> *From: *cp... at googlegroups.com <cp... at googlegroups.com> on behalf of
>>> Nicholas Winner <nwi... at berkeley.edu>
>>> *Date: *Thursday, 10 November 2022 at 18:54
>>> *To: *cp2k <cp... at googlegroups.com>
>>> *Subject: *[CP2K:18011] Thoughts on the instability of HSE06+ADMM
>>>
>>> I'm running GaAs calculation using HSE06 and ADMM basis sets, and
>>> running into some issues.
>>>
>>>
>>>
>>> I'm trying to initialize with a non hybrid calculation and use
>>> SCREEN_ON_INITIAL_P. This is because at 128 atoms (minimum size required
>>> for the HSE06 pot.), the memory requirements get to be pretty large even
>>> compared to other systems.
>>>
>>>
>>>
>>> Problem is that while initializing with PBE may sometimes lead to a
>>> tighter screening threshold, initializing with PBE in this case invariable
>>> leads to a destabilized SCF, with no trend to convergence. I thought that,
>>> since PBE leads to very low band gap, it might be too poor of an initial
>>> guess. So I tried initializing it with (a) r2scan, (b) LDA+OPTX, and (c)
>>> LDA+TB09. These each give better estimates of the band gap and so I would
>>> expect them to provide better initial density matrices, but all of them
>>> failed to create a stable calculation.
>>>
>>>
>>>
>>> Are there any guesses for what could be done and why this is happening?
>>> Does GaAs require no initial guess? That would require some pretty large
>>> memory nodes, but is theoretically doable I guess.
>>>
>>>
>>>
>>> Calculation details:
>>>
>>>
>>>
>>> &QS
>>>
>>> METHOD GPW
>>>
>>> EPS_DEFAULT 1e-12
>>>
>>> EXTRAPOLATION ASPC
>>>
>>> EPS_PGF_ORB 1e-16
>>>
>>> &END QS
>>>
>>>
>>>
>>> ! Parameters needed to perform an SCF run.
>>>
>>> &SCF
>>>
>>> MAX_SCF 20 ! Max number of steps for an inner
>>> SCF loop
>>>
>>> EPS_SCF 1e-06 ! Convergence threshold for SCF
>>>
>>> SCF_GUESS RESTART ! How to initialize the
>>> density matrix
>>>
>>> MAX_ITER_LUMO 400 ! Iterations for solving for
>>> unoccupied levels when running OT
>>>
>>>
>>>
>>> &OT
>>>
>>> MINIMIZER diis
>>>
>>> PRECONDITIONER FULL_SINGLE_INVERSE
>>>
>>> ENERGY_GAP -1
>>>
>>> ALGORITHM IRAC
>>>
>>> LINESEARCH 2PNT
>>>
>>> ROTATION True
>>>
>>> OCCUPATION_PRECONDITIONER False
>>>
>>> &END OT
>>>
>>> &OUTER_SCF
>>>
>>> MAX_SCF 20
>>>
>>> EPS_SCF 1e-06
>>>
>>> &END OUTER_SCF
>>>
>>> &END SCF
>>>
>>> &MGRID
>>>
>>> CUTOFF 350.0
>>>
>>> REL_CUTOFF 50
>>>
>>> NGRIDS 5
>>>
>>> PROGRESSION_FACTOR 3
>>>
>>> &XC
>>>
>>> &XC_FUNCTIONAL
>>>
>>> &PBE
>>>
>>> PARAMETRIZATION ORIG
>>>
>>> SCALE_C 1
>>>
>>> SCALE_X 0
>>>
>>> &END PBE
>>>
>>> &XWPBE
>>>
>>> omega .11
>>>
>>> scale_x -.25
>>>
>>> scale_x0 1
>>>
>>>
>>>
>>> &end
>>>
>>> &END XC_FUNCTIONAL
>>>
>>> &HF
>>>
>>> FRACTION 0.25
>>>
>>> &SCREENING
>>>
>>> EPS_SCHWARZ 1e-07
>>>
>>> EPS_SCHWARZ_FORCES 1e-07
>>>
>>> SCREEN_ON_INITIAL_P T
>>>
>>> SCREEN_P_FORCES False
>>>
>>> &END SCREENING
>>>
>>> &INTERACTION_POTENTIAL
>>>
>>> CUTOFF_RADIUS 6.63
>>>
>>> T_C_G_DATA t_c_g.dat
>>>
>>> omega .11
>>>
>>> POTENTIAL_TYPE shortrange
>>>
>>> &END INTERACTION_POTENTIAL
>>>
>>> &LOAD_BALANCE
>>>
>>> RANDOMIZE True
>>>
>>> &END LOAD_BALANCE
>>>
>>> &MEMORY
>>>
>>> EPS_STORAGE_SCALING 0.1
>>>
>>> MAX_MEMORY 2000
>>>
>>> &END MEMORY
>>>
>>> &END HF
>>>
>>> &END XC
>>>
>>> --
>>> 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/62f09ab5-01d6-497e-8c27-f11b2693756dn%40googlegroups.com
>>> <https://groups.google.com/d/msgid/cp2k/62f09ab5-01d6-497e-8c27-f11b2693756dn%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/6f7bfa9b-6e05-4a6c-bdfd-657b41fcb28fn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20221116/a3bfbcff/attachment-0001.htm>
More information about the CP2K-user
mailing list