About the improvement in scf-convergence.

Top Page > Browsing

About the improvement in scf-convergence.

Date: 2022/06/20 15:39
Name: Takuma Takeda: Dear developers,

Let me ask about convergence about scf calculation.

Now I am conducting a group of calculation of transition metal oxides.
My purpose is to find a stable spin configuration under DFT+U.
Because this system includes some inequivalent sites, I vary Hubbard U term and initial spin ratio (e.g. UP=7.0,DOWN=5.0).
I show one of the example codes below;
#Input--------------------------------------------------------------------------------------------
System.CurrrentDirectory ./
System.Name test
level.of.stdout 1
level.of.fileout 1

Species.Number 4
<Definition.of.Atomic.Species #precise
O O6.0-s3p2d2 O_PBE19
Ti1 Ti7.0-s3p2d2f1 Ti_PBE19
Ti2 Ti7.0-s3p2d2f1 Ti_PBE19
Ti3 Ti7.0-s3p2d2f1 Ti_PBE19
Definition.of.Atomic.Species>

<Hubbard.U.values #eV,UJ2206_002.
O 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 5.6 1d 0.0 2d 0.0
Ti1 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 0.0 1d 3.0 2d 0.0 1f 0.0
Ti2 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 0.0 1d 4.0 2d 0.0 1f 0.0
Ti3 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 0.0 1d 3.0 2d 0.0 1f 0.0
Hubbard.U.values>
#<Hund.J.values #eV
#O 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 0.9 1d 0.0
#Ti 1s 0.0 2s 0.0 3s 0.0 1p 0.0 2p 0.0 1d 0.3 2d 0.0 1f 0.0
##Hund.J.values>

Atoms.Number 16
Atoms.SpeciesAndCoordinates.Unit FRAC
<Atoms.SpeciesAndCoordinates
1 Ti1 0.62755594911239 0.37244227356788 0.05128509798425 7.0 5.0 on
2 Ti1 0.37244002387390 0.62755813208614 0.94871451619810 5.0 7.0 on
3 Ti2 0.30015438960552 0.69984577734319 0.25000567076007 6.0 6.0 on
4 Ti2 0.69982292734663 0.30017704311593 0.75001630103353 6.0 6.0 on
5 Ti3 0.63300963153220 0.36699019243762 0.42072222184749 5.0 7.0 on
6 Ti3 0.36700579653961 0.63299369254725 0.57926923430272 7.0 5.0 on
7 O 0.44368965806196 0.55630956695698 0.38831312812443 3.0 3.0 off
8 O 0.55631229243677 0.44368760870581 0.61168327080359 3.0 3.0 off
9 O 0.17685305480877 0.82314653853640 0.05818001886912 3.0 3.0 off
10 O 0.82315514084934 0.17684443404245 0.94182181326018 3.0 3.0 off
11 O 0.75237720115747 0.24762263029294 0.23813423101361 3.0 3.0 off
12 O 0.24763440288155 0.75236544603376 0.76184934780144 3.0 3.0 off
13 O 0.54530253922425 0.45470076005253 0.86404120525012 3.0 3.0 off
14 O 0.45467962978551 0.54532394310718 0.13595085769106 3.0 3.0 off
15 O 0.18457704018948 0.81542255220195 0.43684391822112 3.0 3.0 off
16 O 0.81543957632688 0.18455995339279 0.56318074740957 3.0 3.0 off
Atoms.SpeciesAndCoordinates>

Atoms.UnitVectors.Unit Ang
<Atoms.UnitVectors
4.918900000000000 1.893370000000000 0.000000000000000
-4.918900000000000 1.893370000000000 0.000000000000000
-0.205603881490000 0.000000000000000 9.370202559488000
Atoms.UnitVectors>

scf.XcType GGA-PBE
scf.SpinPolarization on
scf.Hubbard.U on # on|off, default=off
scf.DFTU.Type 1 # 1:Simplified(Dudarev)|2:General, default=1
scf.dc.Type sFLL
scf.ElectronicTemperature 300.0
scf.energycutoff 500
scf.maxIter 100
scf.EigenvalueSolver band
scf.Kgrid 7 7 3
scf.Mixing.Type rmm-diish
scf.Init.Mixing.Weight 0.0010
scf.Min.Mixing.Weight 0.0001
scf.Max.Mixing.Weight 0.3000
scf.Mixing.History 40
scf.Mixing.StartPulay 60
scf.Mixing.EveryPulay 1
scf.criterion 3.67e-6
scf.stress.tensor on
scf.restart on

MD.Type nomd
MD.Opt.DIIS.History 3 # default=3
MD.Opt.StartDIIS 10 # default=5
MD.Opt.EveryDIIS 200 # default=200
MD.maxIter 1
MD.TimeStep 1.0
MD.Opt.criterion 7.35e-4
<MD.Fixed.Cell.Vectors
0 0 1
0 0 1
0 1 0
MD.Fixed.Cell.Vectors>

geoopt.restart on

#-------------------------------------------------
Thanks for the implementation of rmm-diish, the convergence efficiency in the most case of my calculations have dramatically improved.
However, still in some cases scf calculation did not converged well.
In such cases, NormRD stacked around NormRD=0.01~1 order, even after several hundreds of scf-calculation.
I varied "scf.Max.Mixing.Weight" from 0.0030~0.30, and increased the step of scf.Mixing.StartPulay higher than 100, but these effort
seemed not contributed to solve.
Also, I have tried by changing to rmm-diis, didn't improved, neither.
*)Because my calculations are conducted under DFT+U, I have not tried rmm-diisk nor rmm-diisv yet (this system has small bandgap 0~1
eV).
When I changed to Kerker, NormRD increased more than 10, seemed not to be improved.

So I want to know these three questions.
(1)First question is whether to be the (general) way to improve these stacked calculations.
Yet, I think some of my conditions that resulting bad convergence are physically unreasonable, and the unreasonable conditions might
caused to the bad convergence.

(2)About Kerker calculation.
As my understanding, in the case scf.Mixing.Type=rmm-diis, diisk, and diish condition, they start with Kerker until the step of
scf.Mixing.Startpulay, and during scf.Mixing.EveryPulay works(in the case not equal 1.)
However, the trend of NormRD was different with the case of scf.Mixing.type=Kerker and scf.Mixing.type=rmm-diis,diisk,diish.
I want to know the reason or if there is misunderstanding, I ask to be pointed out.

(3)About the function of scf.Mixing.EveryPulay.
This keyword is included on "Relevant keyword" in "General" section (http://www.openmx-square.org/openmx_man3.9/node40.html) but explained only for rmm-diisk, rmm-diisv and Kerker in the section .
Is it effective to set a value of scf.Mixing.EveryPulay in the case rmm-diis or rmm-diish?
And how does it work if scf.Mixing EveryPulay not as 1, for example "scf.Mixing EveryPulay=5 in the case of rmm-diis or rmm-diish?

I have one more question about Kerker value, but because it is long, let me made another thread.
I apologize for having a lot of questions in one time.
Your Sincerely,

Page: [1]

Re: About the improvement in scf-convergence. ( No.1 )

Date: 2022/07/04 21:53
Name: T. Ozaki

Hi,

With the following parameters:

scf.ElectronicTemperature 700.0
scf.Init.Mixing.Weight 0.0010
scf.Min.Mixing.Weight 0.0001
scf.Max.Mixing.Weight 0.3000
scf.Mixing.History 40
scf.Mixing.StartPulay 30
scf.criterion 1.0e-6

I got the convergent result as

SCF= 1 NormRD= 1.000000000000 Uele= -129.185360080418
SCF= 2 NormRD= 2.141408015849 Uele= -129.183852597014
SCF= 3 NormRD= 2.137171185519 Uele= -128.817009573648
SCF= 4 NormRD= 1.510517706784 Uele= -128.252082436698
SCF= 5 NormRD= 1.082267801352 Uele= -127.459448646141
SCF= 6 NormRD= 0.640037071454 Uele= -126.730339245227
SCF= 7 NormRD= 0.820990317184 Uele= -126.541190619323
SCF= 8 NormRD= 0.380953013210 Uele= -126.376278669652
SCF= 9 NormRD= 0.305294677197 Uele= -126.096754067255
SCF= 10 NormRD= 0.694027075703 Uele= -126.077832643768
SCF= 11 NormRD= 0.289802445133 Uele= -126.054636377625
SCF= 12 NormRD= 0.149497447912 Uele= -126.024093471622
SCF= 13 NormRD= 0.124556794496 Uele= -125.935652523133
SCF= 14 NormRD= 0.094595647859 Uele= -125.878648466012
SCF= 15 NormRD= 0.068769272061 Uele= -125.853248035773
SCF= 16 NormRD= 0.050468837040 Uele= -125.843521581809
SCF= 17 NormRD= 0.037599540455 Uele= -125.841311125921
SCF= 18 NormRD= 0.032276822714 Uele= -125.843739577750
SCF= 19 NormRD= 0.051470249132 Uele= -125.843905640081
SCF= 20 NormRD= 0.019134679948 Uele= -125.844996454881
SCF= 21 NormRD= 0.017304845868 Uele= -125.848156422113
SCF= 22 NormRD= 0.013474289239 Uele= -125.851561376962
SCF= 23 NormRD= 0.013028817913 Uele= -125.854061814166
SCF= 24 NormRD= 0.027607410476 Uele= -125.854761690793
SCF= 25 NormRD= 0.010822037798 Uele= -125.855369190272
SCF= 26 NormRD= 0.007115653626 Uele= -125.856283335197
SCF= 27 NormRD= 0.006292922831 Uele= -125.858452185020
SCF= 28 NormRD= 0.005436082031 Uele= -125.860326664490
SCF= 29 NormRD= 0.005869321488 Uele= -125.860865093893
SCF= 30 NormRD= 0.005798516413 Uele= -125.868370468135
SCF= 31 NormRD= 0.032234648219 Uele= -125.869166494027
SCF= 32 NormRD= 0.027360702868 Uele= -125.869125519817
SCF= 33 NormRD= 0.026023680417 Uele= -125.869067425757
SCF= 34 NormRD= 0.024222069147 Uele= -125.868700015417
SCF= 35 NormRD= 0.023544874640 Uele= -125.868385982196
SCF= 36 NormRD= 0.023815337917 Uele= -125.868091249032
SCF= 37 NormRD= 0.024239765508 Uele= -125.867896787196
SCF= 38 NormRD= 0.024608354996 Uele= -125.867851015428
SCF= 39 NormRD= 0.025141781869 Uele= -125.867844294952
SCF= 40 NormRD= 0.025754468534 Uele= -125.867937940857
SCF= 41 NormRD= 0.026657053617 Uele= -125.868076247504
SCF= 42 NormRD= 0.027295295439 Uele= -125.868112971943
SCF= 43 NormRD= 0.027490369374 Uele= -125.868131300559
SCF= 44 NormRD= 0.027353034383 Uele= -125.867350833305
SCF= 45 NormRD= 0.022955799979 Uele= -125.867701121704
SCF= 46 NormRD= 0.022642870472 Uele= -125.868449386114
SCF= 47 NormRD= 0.026022903124 Uele= -125.868835058559
SCF= 48 NormRD= 0.028157443651 Uele= -125.871983399464
SCF= 49 NormRD= 0.037070925124 Uele= -125.870938520917
SCF= 50 NormRD= 0.033060037368 Uele= -125.873648754570
SCF= 51 NormRD= 0.042033478863 Uele= -125.871782863239
SCF= 52 NormRD= 0.041749071589 Uele= -125.872489358361
SCF= 53 NormRD= 0.034043353742 Uele= -125.873242875537
SCF= 54 NormRD= 0.039099491821 Uele= -125.873532636373
SCF= 55 NormRD= 0.036128399158 Uele= -125.873443020629
SCF= 56 NormRD= 0.031404630303 Uele= -125.872580124037
SCF= 57 NormRD= 0.015383030580 Uele= -125.873980868886
SCF= 58 NormRD= 0.009427884488 Uele= -125.874945309829
SCF= 59 NormRD= 0.005139059443 Uele= -125.875403715175
SCF= 60 NormRD= 0.005698482480 Uele= -125.873809516003
SCF= 61 NormRD= 0.004402542304 Uele= -125.873723497960
SCF= 62 NormRD= 0.003328906607 Uele= -125.873469155607
SCF= 63 NormRD= 0.002160963449 Uele= -125.873322788950
SCF= 64 NormRD= 0.001665454737 Uele= -125.873266017446
SCF= 65 NormRD= 0.001790838946 Uele= -125.873171065854
SCF= 66 NormRD= 0.001280832292 Uele= -125.873316813346
SCF= 67 NormRD= 0.000817457708 Uele= -125.873217032685
SCF= 68 NormRD= 0.000417286941 Uele= -125.873065118996
SCF= 69 NormRD= 0.000368795077 Uele= -125.872937898928
SCF= 70 NormRD= 0.000377620267 Uele= -125.872955812937
SCF= 71 NormRD= 0.000523612418 Uele= -125.872988783597
SCF= 72 NormRD= 0.000621164342 Uele= -125.872963800310
SCF= 73 NormRD= 0.000394822708 Uele= -125.872974619410
SCF= 74 NormRD= 0.000194721784 Uele= -125.873021765938
SCF= 75 NormRD= 0.000088578604 Uele= -125.873024042385
SCF= 76 NormRD= 0.000037980052 Uele= -125.873017163795
SCF= 77 NormRD= 0.000041607891 Uele= -125.873035423871
SCF= 78 NormRD= 0.000031689043 Uele= -125.873039386800
SCF= 79 NormRD= 0.000019301166 Uele= -125.873041928404
SCF= 80 NormRD= 0.000017076482 Uele= -125.873041057668

I wonder that the increase of the electronic temperature improved the convergence.
Anyway, let me reply to your questions:

> (1)First question is whether to be the (general) way to improve these stacked calculations.
> Yet, I think some of my conditions that resulting bad convergence are physically unreasonable,
> and the unreasonable conditions might caused to the bad convergence.

Difficult cases in getting the SCF convergence are related to the electronic states.
For example, one can imagine the following cases:
(i) If nealy degenerate states exist around the Fermi level, the occupations change every SCF step,
which might be your case.
(ii) If highly locaized spin states exist together with delocalized metallic electrons around the Fermi level,
the occupations of both the states change every SCF step, causing the SCF instability.
(iii) If the spin polarized state are energetically competitive with the spin non-polized state, the SCF convergence
cannot be achieved easily, which happens the case of Pd surface. Note that Pd is located on the boarder of the instability
determined by the Stoner condition.
(iv) The lengths of cell vectors are quite different from each other, resulting in a large difference of eigenvalues of
response function, and a bad SCF convergence. 　

Also, it should be noticed that if the geometry optimization is performed, the full SCF convergence is not required for the initial
stage of the geometry optimization. Though the initial structure may have an unstable eletronic state, the SCF convergence
tends to be improved as the geometry optimization proceeds. So, one can relax the threshold for the SCF convergence in the
beginning of the geometry optimization, and after getting an almost convergent geomertical structure, one can make the
threshold severe. This is a widely used technique.

> (2)About Kerker calculation.
> As my understanding, in the case scf.Mixing.Type=rmm-diis, diisk, and diish condition, they start with Kerker until the step of
> scf.Mixing.Startpulay, and during scf.Mixing.EveryPulay works(in the case not equal 1.)
> However, the trend of NormRD was different with the case of scf.Mixing.type=Kerker and scf.Mixing.type=rmm-diis,diisk,diish.
> I want to know the reason or if there is misunderstanding, I ask to be pointed out.

Before starting with the Pulay type mixings such as rmm-diis, diisk, diisv, and diish, the simple mixing is performed.
The Kerker type mixing is the simple mixing in k-space. So, diisk and diisv are relevant to the Kerker type mixing.
Even in these cases, diisk mixes the charge density, while diisv mixes the potential. So, NormRD cannot be equivalent to each other.
In case of diis and diish, the mixing is performed in real space, but diis mixes the density matrix, while diish mixes the Hamiltonian.
So, NormRD cannot be equivalent to each other as well.

> (3)About the function of scf.Mixing.EveryPulay.
> This keyword is included on "Relevant keyword" in "General" section (http://www.openmx-square.org/openmx_man3.9/node40.html)
> but explained only for rmm-diisk, rmm-diisv and Kerker in the section .
> Is it effective to set a value of scf.Mixing.EveryPulay in the case rmm-diis or rmm-diish?
> And how does it work if scf.Mixing EveryPulay not as 1, for example "scf.Mixing EveryPulay=5 in the case of rmm-diis or rmm-diish?

scf.Mixing EveryPulay is valid for only diisk and diisv.
I wonder that the combination of the other keywords replaces the role of scf.Mixing EveryPulay.

Regards,

TO

Re: About the improvement in scf-convergence. ( No.2 )

Date: 2022/06/27 15:37
Name: Takuma Takeda <takuma.takeda.chem@gmail.com>

Dear Prof. Ozaki,

I appreciate for your prompt and helpful advises, and answers against my question.
Also, I apologize being late for my reply.

(1)First of all, I tried under the condition with increased electron temperature 700 K, and succeeded in the SCF calculation with required convergence.
Moreover, most of the other condition was also showed improved convergence.

#Log#
SCF= 1 NormRD= 1.000000000000 Uele= -129.185360080420
SCF= 2 NormRD= 2.141408016004 Uele= -129.183852597017
SCF= 3 NormRD= 2.137171185561 Uele= -128.817009583031
SCF= 4 NormRD= 1.510517706055 Uele= -128.252082457524
SCF= 5 NormRD= 1.082267818281 Uele= -127.459448652965
SCF= 6 NormRD= 0.640037034834 Uele= -126.730339295877
SCF= 7 NormRD= 0.820990085074 Uele= -126.541190630182
SCF= 8 NormRD= 0.380953124451 Uele= -126.376278598107
SCF= 9 NormRD= 0.305294917420 Uele= -126.096754026002
SCF= 10 NormRD= 0.694028171943 Uele= -126.077832633490
SCF= 11 NormRD= 0.289803272850 Uele= -126.054636393872
SCF= 12 NormRD= 0.149497617244 Uele= -126.024093553243
SCF= 13 NormRD= 0.124556819010 Uele= -125.935653125970
SCF= 14 NormRD= 0.094595855654 Uele= -125.878648753765
SCF= 15 NormRD= 0.068769395795 Uele= -125.853248195997
SCF= 16 NormRD= 0.050468868140 Uele= -125.843521539236
SCF= 17 NormRD= 0.037598412900 Uele= -125.841311384976
SCF= 18 NormRD= 0.032267161845 Uele= -125.843738827033
SCF= 19 NormRD= 0.051420978322 Uele= -125.843905871441
SCF= 20 NormRD= 0.019141862837 Uele= -125.844998949572
SCF= 21 NormRD= 0.017302838573 Uele= -125.848158049093
SCF= 22 NormRD= 0.013485016965 Uele= -125.851565649502
SCF= 23 NormRD= 0.013130957481 Uele= -125.854057763364
SCF= 24 NormRD= 0.028040241384 Uele= -125.854752866672
SCF= 25 NormRD= 0.011096901787 Uele= -125.855358974912
SCF= 26 NormRD= 0.007150896577 Uele= -125.856241040106
SCF= 27 NormRD= 0.006321321983 Uele= -125.858418267504
SCF= 28 NormRD= 0.005430697192 Uele= -125.860297289167
SCF= 29 NormRD= 0.005781565783 Uele= -125.860849954853
SCF= 30 NormRD= 0.005816038829 Uele= -125.868368441566
SCF= 31 NormRD= 0.032245567263 Uele= -125.869156366145
SCF= 32 NormRD= 0.027265594790 Uele= -125.869080846877
SCF= 33 NormRD= 0.026130461878 Uele= -125.869033110478
SCF= 34 NormRD= 0.024243065850 Uele= -125.868717765973
SCF= 35 NormRD= 0.023490199733 Uele= -125.868378484665
SCF= 36 NormRD= 0.023747549600 Uele= -125.868110080941
SCF= 37 NormRD= 0.024133313423 Uele= -125.867882699001
SCF= 38 NormRD= 0.024549309437 Uele= -125.867830062029
SCF= 39 NormRD= 0.025083537276 Uele= -125.867823997048
SCF= 40 NormRD= 0.025605318107 Uele= -125.867904882735
SCF= 41 NormRD= 0.026605829497 Uele= -125.868043077167
SCF= 42 NormRD= 0.027228521411 Uele= -125.868081297513
SCF= 43 NormRD= 0.027410319716 Uele= -125.868103630494
SCF= 44 NormRD= 0.027270456028 Uele= -125.867329753936
SCF= 45 NormRD= 0.022990338543 Uele= -125.867704025784
SCF= 46 NormRD= 0.022660982402 Uele= -125.868478642075
SCF= 47 NormRD= 0.026124605949 Uele= -125.868830205098
SCF= 48 NormRD= 0.028200172307 Uele= -125.871915179964
SCF= 49 NormRD= 0.036768705009 Uele= -125.870881883141
SCF= 50 NormRD= 0.032915659170 Uele= -125.873543594667
SCF= 51 NormRD= 0.042059618977 Uele= -125.871756194841
SCF= 52 NormRD= 0.041690559488 Uele= -125.872533158209
SCF= 53 NormRD= 0.033736387215 Uele= -125.873322651401
SCF= 54 NormRD= 0.039201038870 Uele= -125.873561643086
SCF= 55 NormRD= 0.036301918191 Uele= -125.873554825835
SCF= 56 NormRD= 0.033170640891 Uele= -125.872710009723
SCF= 57 NormRD= 0.018438680891 Uele= -125.873631073840
SCF= 58 NormRD= 0.011757465825 Uele= -125.874786898814
SCF= 59 NormRD= 0.004867907073 Uele= -125.875460705015
SCF= 60 NormRD= 0.004703292200 Uele= -125.873532969784
SCF= 61 NormRD= 0.004103152169 Uele= -125.873740457736
SCF= 62 NormRD= 0.003247939613 Uele= -125.873444784216
SCF= 63 NormRD= 0.002109081983 Uele= -125.873278659135
SCF= 64 NormRD= 0.001742669713 Uele= -125.873239577420
SCF= 65 NormRD= 0.002063253416 Uele= -125.873021746347
SCF= 66 NormRD= 0.001561071845 Uele= -125.873116546703
SCF= 67 NormRD= 0.001558181025 Uele= -125.872957652885
SCF= 68 NormRD= 0.001294870926 Uele= -125.872961451770
SCF= 69 NormRD= 0.001071443656 Uele= -125.872903349402
SCF= 70 NormRD= 0.001318289003 Uele= -125.872887138631
SCF= 71 NormRD= 0.001336167354 Uele= -125.872940380572
SCF= 72 NormRD= 0.001274340241 Uele= -125.872987513661
SCF= 73 NormRD= 0.001283975901 Uele= -125.872978685809
SCF= 74 NormRD= 0.001323147844 Uele= -125.873085467210
SCF= 75 NormRD= 0.001140872770 Uele= -125.873042168280
SCF= 76 NormRD= 0.000570608539 Uele= -125.873294702665
SCF= 77 NormRD= 0.000280562442 Uele= -125.873110524195
SCF= 78 NormRD= 0.000163563440 Uele= -125.873050364663
SCF= 79 NormRD= 0.000052662424 Uele= -125.873059317837
SCF= 80 NormRD= 0.000036452020 Uele= -125.873044175932
SCF= 81 NormRD= 0.000021463969 Uele= -125.873042456936
#####
The steps required to complete the SCF calculation was slight different from yours, but it is just because of the difference of computational environment, not so big issue, I think.
(I compared with AMD-Ryzen+MateriAppsLive!3.2+OpenMX3.9.2 and Intel-Corei5+MateriAppsLive!3.2+OpenMX3.9.2, and these two conditions showed slight different trend in NormRD and Uele.)

(2)About the questions for SCF calculation including Kerker factor in the next thread, the answers were so clear that I got it.
Surely, I started with the structure, which was relaxed under another condition. By varying U and spin configuration, extra atomic forces as high as 0.1 eV/Å emerged.
(Moreover, I set the U value for oxygen by mistake, it was 3p-orbital in real space, not 2p-orbitals. I should have set for 2p.)
I thought that I cannot relax the structures because their SCF had not successfully converged.
I will try the relaxation for each condition, by applying high electronic temperature and untightening SCF in the several first steps.

Let me ask additional questions about SCF.ElectronicTemperature and SCF convergence for structural relaxation.
(Q1)As my understanding, SCF.ElectronicTemperature indicates the degree for the kind of "smearing" in Fermi-Dirac dispersion.
Thus, I thought that it is suitable to set at most room temperature to calculate basis state.
I found some reports applied higher electronic temperature than 1000 K, to take account in the contribution of the higher energy level (e.g. Ab-initio molecular dynamics in DOI: 10.1126/science.aau3873, see supporting information, which was performed with VASP though,) but they are seemed to be special cases.
I wonder how degree of the maximum value is normal for usual SCF calculation.

(Q2)Is there limit for the degree of untightening SCF criterion in the beginning steps of structural relaxation?
Normally SCF criterion requires less than 1E-6 eV, so I think 1E-4 ~ 1E-5 eV criterion seemed to be suitable.

Your Sincerely,

Re: About the improvement in scf-convergence. ( No.3 )

Date: 2022/06/29 21:53
Name: T. Ozaki

Hi,

>(Q1)As my understanding, SCF.ElectronicTemperature indicates the degree for the kind of "smearing" in Fermi-Dirac dispersion.
>Thus, I thought that it is suitable to set at most room temperature to calculate basis state.
>I found some reports applied higher electronic temperature than 1000 K, to take account in the contribution of the higher energy level (e.g.
>Ab-initio molecular dynamics in DOI: 10.1126/science.aau3873, see supporting information, which was performed with VASP though,) but they are
>seemed to be special cases.
>I wonder how degree of the maximum value is normal for usual SCF calculation.

For gapped systems, results such as energetics and optimized structures are less sensitive to electronic temperature unless it is beyond 1000K.
Even to metals, the statement is true in most cases. However, in much higher temperature the energetics and optimized structures are influenced,
since the occupations increase in untibonding states. To figure out how results are changed depending on electronic temperature, you can perform
benchmark calculations for systems which are similar to your systems but easy to handle.

>(Q2)Is there limit for the degree of untightening SCF criterion in the beginning steps of structural relaxation?
>Normally SCF criterion requires less than 1E-6 eV, so I think 1E-4 ~ 1E-5 eV criterion seemed to be suitable.

If the SCF convergence reaches to 1E-4 ~ 1E-5 eV, the error in forces is acceptable in most cases.
This can also be checked by performing some benchmark calculations.

Regards,

TO

Re: About the improvement in scf-convergence. ( No.4 )

Date: 2022/06/30 21:43
Name: Takuma Takeda

Dear Prof. Ozaki,

I again appreciate for your kind reply.

It's meaningful for me to know about electronic temperature and convergence in structural relaxation.
These days I am conducting tens of conditions with the settings obtained in this thread, and they are perfectly completed, including structural relaxation. Your advises were so helpful.

Sincerely yours,

Takeda

Page: [1]