[QE-users] non-convergency of DFPT+U calculation
Timrov Iurii
iurii.timrov at epfl.ch
Wed Mar 4 17:58:22 CET 2020
Dear Merlin,
Concerning phonons from DFPT+U: please make sure that you have an insulator (check the HOMO and LUMO energies in the output of the second SCF run, check bands and DOS). From my experience with DFPT+U for metals, when you use finite U it diverges often (only for small U of 1-2 eV it was converging e.g. in bulk Ni) - even if I perform a frozen-phonon calculations with supercells for metals with finite U there are convergence issues, which suggests that it is a physical problem. This has to be investigated more.
FeO is a complex system, it is metallic even when you apply U in a "standard" way (to open a gap one should use starting_ns_eigenvalue, please check tutorials on this topic). So since FeO is metallic, and for metals DFPT+U is diverging in many cases (as said above), I do not know what is the solution. If you open a gap in FeO, I presume phonons should be OK with finite U.
In general: always make sure that you indeed have an insulator (check the band structure and DOS).
Concerning the calculation of U from DFPT: "Some work, some do not converge, some diverge." - this does not tell me much, I need to see all input and output files to make suggestions.
Greetings,
Iurii
--
Dr. Iurii Timrov
Postdoctoral Researcher
STI - IMX - THEOS and NCCR - MARVEL
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne, Switzerland
+41 21 69 34 881
http://people.epfl.ch/265334
________________________________
From: users <users-bounces at lists.quantum-espresso.org> on behalf of Merlin Meheut <merlin.meheut at gmail.com>
Sent: Wednesday, March 4, 2020 5:32:47 PM
To: Quantum ESPRESSO users Forum
Subject: Re: [QE-users] non-convergency of DFPT+U calculation
Dear Iurii,
I learned a lot with your guidance, thank you so much! Unfortunately, I am afraid that none of your propositions (use SSSP pseudopotentials, change density cutoff) did solve the matter (1). To further explore this question, I also did DFPT+U phonon calculation on FeO (2), with an input almost identical to your example02 on NiO, and there was a divergence for the second representation, very similar to what I have for Fe2SiO4. At last (3), I tried to compute the +U value based on your example02 as well. For fayalite, I succeeded to compute U for 1x1x1 q-point grid, but 2x1x2 and 2x2x2 grids failed. For FeO, the situation is more complex, depending on the k-point grid and on the q-point grid, sometimes it works, sometimes it does not converge, sometimes it crashes. But generally for large grids it is not possible to reach the end of the calculation. Below are more detailed descriptions of these problems:
The procedures were tested for pseudopotentials
O.pz-n-rrkjus_psl.0.1.UPF + Fe.pz-spn-rrkjus_psl.0.2.1.UPF + Si.pz-n-rrkjus_psl.0.1.UPF
Fe.pbesol-spn-kjpaw_psl.0.2.1.UPF + Si.pbesol-n-rrkjus_psl.1.0.0.UPF + O.pbesol-n-kjpaw_psl.0.1.UPF
and the formerly described pseudopotentials
(1) For DFPT+U phonon calculations on Fayalite, I followed your suggestion to adopt the approach described in example 02, with a first scf calculation in the metallic state:
&system
(...)
nspin=2,
occupations='smearing',
degauss=0.001,
smearing= 'gauss',
starting_magnetization(1)=0.5,
starting_magnetization(2)=-0.5,
lda_plus_u = .true.,
lda_plus_u_kind = 0,
U_projection_type = 'ortho-atomic',
Hubbard_U(1)=2.5,
Hubbard_U(2)=2.5,
/&end
&electrons
electron_maxstep = 200,
conv_thr = 1.d-10,
mixing_beta = 0.4,
/&end
(...)
K_POINTS {automatic}
2 2 2 1 1 1
and a second calculation as insulating material:
&system
ibrav =8,
celldm(1)=9.107535,
celldm(2)=2.17425,
celldm(3)=1.263056,
nat =28,
ntyp = 4,
ecutwfc =80.0,
ecutrho = 320.0,
nbnd=${nstates},
nspin=2,
occupations='fixed',
tot_magnetization= 0,
lda_plus_u = .true.,
lda_plus_u_kind = 0,
U_projection_type = 'atomic',
Hubbard_U(1)=2.5,
Hubbard_U(2)=2.5,
/&end
&electrons
electron_maxstep = 200,
conv_thr = 1.d-15,
mixing_mode = 'plain',
mixing_beta = 0.4,
startingpot = 'file',
startingwfc = 'file',
/&end
followed by the phonon calculation:
Liz m1
&inputph
alpha_mix(1) = 0.3,
tr2_ph = 1.0D-17,
prefix='FAYA2',
fildyn='mat.$PREFIX',
lraman=.false.,
epsil =.false.,
trans =.true.,
zue = .false.,
zeu= .false.,
start_irr=23,
last_irr=23,
niter_ph=20,
outdir ='./',
/&end
0.0 0.0 0.0
The procedure was tested for pseudopotentials
O.pz-n-rrkjus_psl.0.1.UPF + Fe.pz-spn-rrkjus_psl.0.2.1.UPF + Si.pz-n-rrkjus_psl.0.1.UPF
Fe.pbesol-spn-kjpaw_psl.0.2.1.UPF + Si.pbesol-n-rrkjus_psl.1.0.0.UPF + O.pbesol-n-kjpaw_psl.0.1.UPF
and the formerly described
I also changed cutoffs as you suggested. There is no difference that I can see in the result.
(2) For DFPT+U phonon calculations on FeO, I adopted the same approach, with two scf calculations very similar to the one shown in HP example02, except for U_projection_type = 'atomic', instead of 'ortho-atomic' for the second scf . The following phonon input looks like this:
Liz m1
&inputph
alpha_mix(1) = 0.3,
tr2_ph = 1.0D-15,
(...)
lraman=.false.,
epsil =.false.,
trans =.true.,
zue = .false.,
zeu= .false.,
start_irr=2,
last_irr=2,
niter_ph=20,
/&end
0.0 0.0 0.0
Here also, we obtain a very clear divergence, whatever the Fe and O pseudopotentials considered above.
(3) Then I tried to compute the value of the +U parameter, based on your HP example02. I did it for fayalite and for FeO (in which case the input is essentially the same than example02, except for the pseudopotentials). Following your article (Timrov et al 2018, PRB 98, 085127), I tried different values for the k-point and q-point grids . Some work, some do not converge, some diverge. For fayalite, it was only possible to converge for 1x1x1 q-point grid. For FeO as well, large grids tend to be just impossible to complete, as some particular q-points do not converge. Once again, whatever the pseudo or the cutoff. However, these are not exactly the same configurations that fail depending on the pseudopotentials.
If you have any hints that may permit to progress on those issues that would be awesome!
Best regards,
Merlin
--
Merlin Méheut
adresse labo:
GET - OMP - Université Paul Sabatier
14 avenue Edouard Belin
31400 Toulouse
FRANCE
tel: (+33) 5 61 33 26 17
http://www3.obs-mip.fr/get/profils/Meheut_Merlin
Le jeu. 20 févr. 2020 à 10:05, Timrov Iurii <iurii.timrov at epfl.ch<mailto:iurii.timrov at epfl.ch>> a écrit :
Dear Merlin,
By the way, since you are using DFPT+U to compute phonons including the Hubbard U correction, I would like to let you (and others) know that the paper describing DFPT+U has been just published in Physical Review B:
A. Floris, I. Timrov, B. Himmetoglu, N. Marzari, S. de Gironcoli, and M. Cococcioni, "Hubbard-corrected density functional perturbation theory with ultrasoft pseudopotentials", Phys. Rev. B 101, 064305 (2020).
Best regards,
Iurii
--
Dr. Iurii Timrov
Postdoctoral Researcher
STI - IMX - THEOS and NCCR - MARVEL
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne, Switzerland
+41 21 69 34 881
http://people.epfl.ch/265334
________________________________
From: users <users-bounces at lists.quantum-espresso.org<mailto:users-bounces at lists.quantum-espresso.org>> on behalf of Timrov Iurii <iurii.timrov at epfl.ch<mailto:iurii.timrov at epfl.ch>>
Sent: Wednesday, February 19, 2020 10:05:01 AM
To: Quantum ESPRESSO users Forum
Subject: Re: [QE-users] non-convergency of DFPT+U calculation
Dear Merlin,
Total magnetization must be zero for an antiferromagnetic insulator. Hence, some deviation from zero that you obtain is probably a numerical noise (I would check the convergence of the results: reduce conv_thr, increase the size of the k mesh, etc., and relax well the structure).
> If I wanted to be closer to the smearing calculation, I would have to consider nocolin=.true., is that right?
DFPT+U with noncolin=.true. is not implemented.
Greetings,
Iurii
--
Dr. Iurii Timrov
Postdoctoral Researcher
STI - IMX - THEOS and NCCR - MARVEL
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne, Switzerland
+41 21 69 34 881
http://people.epfl.ch/265334
________________________________
From: users <users-bounces at lists.quantum-espresso.org<mailto:users-bounces at lists.quantum-espresso.org>> on behalf of Merlin Meheut <merlin.meheut at gmail.com<mailto:merlin.meheut at gmail.com>>
Sent: Wednesday, February 19, 2020 12:38:08 AM
To: Quantum ESPRESSO users Forum
Subject: Re: [QE-users] non-convergency of DFPT+U calculation
Dear Iurii,
Thank you again for your help! So far, the Example02 seems quite clear to me, although it will take me some time to apply it to my system and check that everything is working. One quick question, though: in my case, the "total magnetization" obtained from the first calculation (with smearing) is non-zero and non-integer: "total magnetization = -0.63 Bohr mag/cell" (which makes sense since I have two different sites). As I understand it, for the second calculation, the variable tot_magnetization should be integer, so I can only take it equal to 0 or 1 here (I tried with 0, I don't know which is better). If I wanted to be closer to the smearing calculation, I would have to consider nocolin=.true., is that right?
Greetings,
Merlin
Le mar. 18 févr. 2020 à 17:32, Timrov Iurii <iurii.timrov at epfl.ch<mailto:iurii.timrov at epfl.ch>> a écrit :
Dear Merlin,
> Does that mean that magnetic insulators cannot be treated by Quantum Espresso?
It is possible to model magnetic insulators in Quantum ESPRESSO. I am using the procedure described in q-e-qe-6.5/HP/examples/example02/README. If the description in README file is not clear enough, please let me know.
Greetings,
Iurii
--
Dr. Iurii Timrov
Postdoctoral Researcher
STI - IMX - THEOS and NCCR - MARVEL
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne, Switzerland
+41 21 69 34 881
http://people.epfl.ch/265334
________________________________
From: users <users-bounces at lists.quantum-espresso.org<mailto:users-bounces at lists.quantum-espresso.org>> on behalf of Merlin Meheut <merlin.meheut at gmail.com<mailto:merlin.meheut at gmail.com>>
Sent: Tuesday, February 18, 2020 5:21:46 PM
To: Quantum ESPRESSO users Forum
Subject: Re: [QE-users] non-convergency of DFPT+U calculation
Hi Iurii,
First thank you very much for this remark, I did not know that one could not use tot_magnetization and starting_magnetization together. Indeed, this is reported in the Documentation INPUT_PW.def.
At this point, however, I don't see how to make an input for an antiferromagnetic insulating material (which is the goal here). Because with this input, the state is non magnetic:
&system
ibrav =8, celldm(1)=9.107535, celldm(2)=2.17425,
celldm(3)=1.263056,
nat =28, ntyp = 4, ecutwfc =80.0, ecutrho = 320.0,
nspin=2, tot_magnetization=0,
lda_plus_u = .true., U_projection_type = 'atomic',
Hubbard_U(1)=2.4, Hubbard_U(2)=2.6, ! valeurs Yu et al 2013 =2.4/2.6
/&end
And if you use this input:
&system
ibrav =8, celldm(1)=9.107535, celldm(2)=2.17425,
celldm(3)=1.263056,
nat =28, ntyp = 4, ecutwfc =${a}.0, ecutrho = ${b}.0,
nspin=2,
starting_magnetization(1)=0.5, starting_magnetization(2)=-0.5,
lda_plus_u = .true., U_projection_type = 'atomic',
Hubbard_U(1)=2.5, Hubbard_U(2)=2.5,
/&end
,you have an error message:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Error in routine iosys (1):
fixed occupations and lsda need tot_magnetization
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
So as you say, using starting_magnetization imposes using smearing. But if I am correct, smearing means that your material is metallic, and therefore you cannot compute dielectric quantities or effective charges with phonon. Actually, in the INPUT_PW.def file, the word "antiferromagnetic" is only present once, in the "starting_magnetization" definitions. Does that mean that magnetic insulators cannot be treated by Quantum Espresso?
Would you have any hint as to how to extract myself from such a dead end?
Thanks again for your help,
Greetings,
Merlin
Le dim. 16 févr. 2020 à 23:11, Timrov Iurii <iurii.timrov at epfl.ch<mailto:iurii.timrov at epfl.ch>> a écrit :
Dear Merlin,
In your input for the PW calculation you have:
> tot_magnetization=0,
> starting_magnetization(1)=0.5, starting_magnetization(2)=-0.5
tot_magnetization and starting_magnetization must not be used together (please check the documentation). Moreover, starting_magnetization must be used with occupations='smearing', but I do not see smearing in your input. Is the ground state metallic or insulating?
How many representations do you have in total? And only #23 diverges?
Not sure this will help, but you may try to use dual = ecutrho/ecutwfc = 12, with ecutwfc=80 Ry.
> Si 27.9769 Si.pz2.UPF
> O 15.9949 O.pz.UPF
I do not know what are these pseudos, and how good/bad are they. You can try to check the SSSP library.
> Fe1 55.9349 Fe.pz-n-rrkjus_psl.1.0.0.UPF
> Fe2 55.9349 Fe.pz-n-rrkjus_psl.1.0.0.UPF
Also for these I would check the SSSP library.
Greetings,
Iurii
--
Dr. Iurii Timrov
Postdoctoral Researcher
STI - IMX - THEOS and NCCR - MARVEL
Swiss Federal Institute of Technology Lausanne (EPFL)
CH-1015 Lausanne, Switzerland
+41 21 69 34 881
http://people.epfl.ch/265334
________________________________
From: users <users-bounces at lists.quantum-espresso.org<mailto:users-bounces at lists.quantum-espresso.org>> on behalf of Merlin Meheut <merlin.meheut at gmail.com<mailto:merlin.meheut at gmail.com>>
Sent: Sunday, February 16, 2020 9:22:57 PM
To: Quantum Espresso users Forum
Subject: [QE-users] non-convergency of DFPT+U calculation
Dear PWSCF users,
I am trying to realize a phonon calculation (at gamma) with +U on fayalite (olivine structure Fe2SiO4) with anti-ferromagnetic configuration , and whether the calculation goes fine without +U, I have one representation (#23) that systematically fails to converge when I add a +U. I tried several values for the +U (0.5, 2.5, 4), several functionals (PZ, PBE), but the result is always the same. I have used QE versions 6.4 and 6.4.1 on two different computing centers . Would you have any idea about the reasons for such a failure? I am specifically worried by one output message: "Atomic wfc used for the DFT+U projector are NOT orthogonalized", but I don't know how to fix that.
Here are examples of my scf and ph inputs, and the failing ph output:
scf:
&control
calculation = 'scf',
(...)
tprnfor = .true.,
tstress = .true.,
/&end
&system
ibrav =0, celldm(1)=9.107535,
nat =28, ntyp = 4, ecutwfc =80.0, ecutrho = 320.0,
nspin=2, tot_magnetization=0,
starting_magnetization(1)=0.5, starting_magnetization(2)=-0.5,
lda_plus_u = .true., U_projection_type = 'atomic',
Hubbard_U(1)=2.4, Hubbard_U(2)=2.6
/&end
&electrons
electron_maxstep = 150,
conv_thr = 1.d-11,
mixing_mode = 'plain',
startingwfc = 'atomic',
mixing_beta = 0.3,
/&end
ATOMIC_SPECIES
Fe1 55.9349 Fe.pz-n-rrkjus_psl.1.0.0.UPF
Fe2 55.9349 Fe.pz-n-rrkjus_psl.1.0.0.UPF
Si 27.9769 Si.pz2.UPF
O 15.9949 O.pz.UPF
(...)
ph input:
&inputph
amass(1)=55.9349,
amass(2)=55.9349,
amass(3)=27.9769,
amass(4)=15.9949,
alpha_mix(1) = 0.5,
! ldisp=.true., nq1=2, nq2=2, nq3=2,
tr2_ph = 1.0D-17,
prefix='FAYA2',
fildyn='mat.$PREFIX',
lraman=.false.,
epsil =.false.,
trans =.true.,
zue = .false.,
zeu= .false.,
start_irr=23,
last_irr=23,
outdir ='./',
/&end
0.0 0.0 0.0
ph output:
Representation # 23 mode # 23
Self-consistent Calculation
iter # 1 total cpu time : 58.7 secs av.it<http://av.it>.: 8.5
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 1.045E-06
iter # 2 total cpu time : 62.9 secs av.it<http://av.it>.: 23.0
thresh= 1.022E-04 alpha_mix = 0.500 |ddv_scf|^2 = 2.668E-04
iter # 3 total cpu time : 66.4 secs av.it<http://av.it>.: 17.5
thresh= 1.633E-03 alpha_mix = 0.500 |ddv_scf|^2 = 3.033E-03
iter # 4 total cpu time : 69.0 secs av.it<http://av.it>.: 11.0
thresh= 5.508E-03 alpha_mix = 0.500 |ddv_scf|^2 = 5.602E-02
iter # 5 total cpu time : 73.0 secs av.it<http://av.it>.: 15.5
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 1.238E+00
iter # 6 total cpu time : 76.7 secs av.it<http://av.it>.: 19.5
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 2.841E+01
iter # 7 total cpu time : 80.9 secs av.it<http://av.it>.: 23.0
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 6.534E+02
iter # 8 total cpu time : 85.3 secs av.it<http://av.it>.: 24.5
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 1.446E+04
iter # 9 total cpu time : 90.1 secs av.it<http://av.it>.: 27.5
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 3.214E+05
iter # 10 total cpu time : 95.5 secs av.it<http://av.it>.: 30.0
thresh= 1.000E-02 alpha_mix = 0.500 |ddv_scf|^2 = 7.200E+06
(.....)
Thank you in advance for your help! Feel free to ask for any further calculation details.
Regards,
--
Merlin Méheut
adresse labo:
GET - OMP - Université Paul Sabatier
14 avenue Edouard Belin
31400 Toulouse
FRANCE
tel: (+33) 5 61 33 26 17
_______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu/quantum-espresso<http://www.max-centre.eu/quantum-espresso>)
users mailing list users at lists.quantum-espresso.org<mailto:users at lists.quantum-espresso.org>
https://lists.quantum-espresso.org/mailman/listinfo/users
--
Merlin Méheut
adresse labo:
GET - OMP - Université Paul Sabatier
14 avenue Edouard Belin
31400 Toulouse
FRANCE
tel: (+33) 5 61 33 26 17
_______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu/quantum-espresso<http://www.max-centre.eu/quantum-espresso>)
users mailing list users at lists.quantum-espresso.org<mailto:users at lists.quantum-espresso.org>
https://lists.quantum-espresso.org/mailman/listinfo/users
--
Merlin Méheut
adresse labo:
GET - OMP - Université Paul Sabatier
14 avenue Edouard Belin
31400 Toulouse
FRANCE
tel: (+33) 5 61 33 26 17
_______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu/quantum-espresso<http://www.max-centre.eu/quantum-espresso>)
users mailing list users at lists.quantum-espresso.org<mailto:users at lists.quantum-espresso.org>
https://lists.quantum-espresso.org/mailman/listinfo/users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20200304/d250f8c3/attachment.html>
More information about the users
mailing list