[QE-users] [QE users] pseudopotential hardness and transferability
Aldo Ugolotti
a.ugolotti at campus.unimib.it
Thu Apr 2 18:16:46 CEST 2020
Dear Nicola,
thanks for your reply. I mentioned the energy criteria for convergence
as the simplest one to verify.
Regarding the pseudos you mentioned (PBE), those available at the
materialscloud portal, I have checked them either in the atomic case and
for a simple system of interest. In the latter case, I found that the
required minimal cutoffs are ~90/225 Ry while in the former case the
results of ld1.x (for Hamiltonian evaluation with Bessel functions) are
consistent with the numbers you reported.
For both N and C pseudos the test on electronic configurations with more
electrons fail as the scf cycle does not converge, both for AE and PS
wavefucntions.
If I understood correctly the works you referred to, the delta parameter
seems much more large-scale (i.e. code to code) than the scope I am
working in, as I am not validating a pseudo generation procedure, (which
I am assuming to be already accurate, given the source) but rather I am
trying to understand the system-dependency of the behavior of the
pseudo. However, I am mistaken, I will be glad to check those papers
again and run some tests.
Bests Regards,
--
Aldo Ugolotti, Ph.D.
Post-doc fellow
Materials Science Dept. U5,
Università degli Studi di Milano-Bicocca
via Cozzi 55,
20125 Milano (MI)
Italy
e-mail: a.ugolotti at campus.unimib.it
On 02/04/20 13:39, Nicola Marzari wrote:
>
>
>
> Dear Aldo,
>
>
> very worthwhile work! Also very hard (pun intended) - pseudopotential
> generation is not an easy task. In particular, you can make things
> softer, but less accurate, and that's not one wants. Also, being at a
> cutoff that converges the total energy is not superinteresting - we
> typically want stresses, forces, et al to be both converged (easy to
> test), and accurate (more difficult, you need to compare with
> all-electron calculations, ideally in some real-life solid state
> systems).
>
> If you go to https://www.materialscloud.org/discover/sssp/ you could
> acquaint yourself with a couple of papers on verification (see section
> "How to cite") and on pseudopotential generation (see section
> "Acknolwedgments").
>
> You could also try out the C and N pseudopotentials from the PBE or
> PBEsol SSSP efficiency library, and see if you really need to have
> larger cutoffs than those suggested (45/360 Ry for ecutwfc and ecutrho
> respectively (i.e. a dual of 8) for the carbon pseudopotential, taken
> as PAW from pslibrary 1.0, and 60/480 Ry for N, generated by me). The
> delta value for the elemental solid is a first measure of how good
> things are with respect to all electron results.
>
> nicola
>
>
>
>
>
> On 02/04/2020 13:15, Aldo Ugolotti wrote:
>> Dear QE users,
>>
>> I am actually working on a system with C and N atoms. Checking the
>> convergence of the total energy for finding the optimal values for
>> the cutoffs (i.e. DE ~ 1mRy), I found that, despite in the atomic
>> case the suggested values (for example the wfc cutoff are ~ 46 Ry for
>> both) are good enough, in a sample of my system which is already
>> relaxed (and whose geometry is in good agreement with reported
>> results) the same convergence check determines a cutoff which is,
>> again for example for the wavefunction, 2 to 3 times larger.
>>
>> As I tried to modify the pseudo to make it softer, I have also run
>> some transferability tests, which I am curious to hear your opinion
>> about. In particular, the tests were running fine for the testing
>> configurations with less electrons (e.g. 2s2 2p1 for C) but there
>> were problems with tests with more electrons (e.g. 2s2 2p3 for C). In
>> those cases the scf cycles did not converge at all, both at AE or PS
>> level.
>>
>> I found the same result with the pseudo US,PAW in the pslibrary of
>> different versions, namely 1.0.0, 0.3.1 and 0.1. I also tried to
>> change the radii, the local potential (adding a 3D empty orbital),
>> the configuration (e.g, Ztot=5.5, Zval 1.5 for C) or the pseudization
>> recipe (TM/RRKJUS).
>>
>> Hence, I got few questions:
>>
>> i) is it really a transferability issue, or do I need "only" to get
>> those scf cycles to converge? how?
>>
>> ii) if the pseudo is not good to represent electronic configurations
>> with more electrons, that would be a viable explanation as to why the
>> cutoffs for a sample systems are so much larger than the atomic cases?
>>
>> Below I am reporting the output for the test of the configtf(2)='2s2
>> 2p3'
>>
>>
>> Message from routine scf:
>> warning: convergence not achieved
>> --------------------------- All-electron run
>> ----------------------------
>>
>> C
>> scalar relativistic calculation
>>
>> atomic number is 6.00
>> dft =SLA PW PBX PBC lsd =0 sic =0 latt =0 beta=0.20 tr2=1.0E-14
>> Exchange-correlation = SLA PW PBX PBC ( 1 4 3 4 0 0)
>> mesh =1073 r(mesh) = 100.30751 a.u. xmin = -7.00 dx = 0.01250
>> 1 Ry = 13.60569193 eV, c = 137.03599966
>>
>> n l nl e(Ry) e(Ha) e(eV)
>> 1 0 1S 1( 2.00) -19.6664 -9.8332 -267.5745
>> 2 0 2S 1( 2.00) -0.6297 -0.3148 -8.5669
>> 2 1 2P 1( 3.00) -0.0290 -0.0145 -0.3951
>>
>> final scf error: 2.4E-01 reached in 201 iterations
>>
>> Etot = -78.638531 Ry, -39.319266 Ha, -1069.931632 eV
>>
>> Ekin = 73.218424 Ry, 36.609212 Ha, 996.187324 eV
>> Encl = -182.081805 Ry, -91.040902 Ha, -2477.348944 eV
>> Eh = 40.989732 Ry, 20.494866 Ha, 557.693668 eV
>> Exc = -10.764883 Ry, -5.382441 Ha, -146.463680 eV
>>
>>
>> normalization and overlap integrals
>>
>> s(1S/1S) = 1.000000 <r> = 0.2707 <r2> = 0.0993 r(max) =
>> 0.1730
>> s(1S/2S) = -0.000112
>> s(2S/2S) = 1.000000 <r> = 1.6236 <r2> = 3.2283 r(max) =
>> 1.2315
>> s(2P/2P) = 1.000000 <r> = 2.1244 <r2> = 6.4268 r(max) =
>> 1.2470
>>
>> ------------------------ End of All-electron run
>> ------------------------
>>
>> Message from routine run_pseudo:
>> Warning: convergence not achieved
>>
>> ---------------------- Testing the pseudopotential
>> ----------------------
>>
>> C
>> scalar relativistic calculation
>>
>> atomic number is 6.00 valence charge is 4.00
>> dft =SLA PW PBX PBC lsd =0 sic =0 latt =0 beta=0.20 tr2=1.0E-14
>> mesh =1073 r(mesh) = 100.30751 xmin = -7.00 dx = 0.01250
>>
>> n l nl e AE (Ry) e PS (Ry) De AE-PS (Ry)
>> 1 0 2S 1( 2.00) -0.62966 -0.17919 -0.45046 !
>> 2 1 2P 1( 3.00) -0.02904 -0.00000 -0.02904 !
>>
>> eps = 3.2E-04 iter =201
>>
>> Etot = -78.638531 Ry, -39.319266 Ha, -1069.931632 eV
>> Etotps = -18.974270 Ry, -9.487135 Ha, -258.158068 eV
>> dEtot_ae = -3.108582 Ry
>> dEtot_ps = -1.208418 Ry, Delta E= -1.900164 Ry
>>
>> Ekin = 10.222924 Ry, 5.111462 Ha, 139.089950 eV
>> Encl = -31.022876 Ry, -15.511438 Ha, -422.087699 eV
>> Ehrt = 12.620743 Ry, 6.310371 Ha, 171.713935 eV
>> Ecxc = -10.795060 Ry, -5.397530 Ha, -146.874254 eV
>> (Ecc = -0.958640 Ry, -0.479320 Ha, -13.042955 eV)
>>
>> ---------------------- End of pseudopotential test
>> ----------------------
>>
>>
>> -------------- Test with a basis set of Bessel functions
>> ----------
>>
>> Box size (a.u.) : 30.0
>>
>> Cutoff (Ry) : 30.0
>> N = 1 N = 2 N = 3
>> E(L=0) = -0.1788 Ry 0.1213 Ry 0.1854 Ry
>> E(L=1) = 0.1263 Ry 0.1949 Ry 0.2715 Ry
>>
>> Cutoff (Ry) : 60.0
>> N = 1 N = 2 N = 3
>> E(L=0) = -0.1789 Ry 0.1213 Ry 0.1854 Ry
>> E(L=1) = 0.1263 Ry 0.1949 Ry 0.2715 Ry
>>
>> Cutoff (Ry) : 90.0
>> N = 1 N = 2 N = 3
>> E(L=0) = -0.1789 Ry 0.1213 Ry 0.1854 Ry
>> E(L=1) = 0.1263 Ry 0.1949 Ry 0.2715 Ry
>>
>> Cutoff (Ry) : 120.0
>> N = 1 N = 2 N = 3
>> E(L=0) = -0.1789 Ry 0.1213 Ry 0.1854 Ry
>> E(L=1) = 0.1263 Ry 0.1948 Ry 0.2715 Ry
>>
>> -------------- End of Bessel function test
>> ------------------------
>>
>>
>> Thank you in advance,
>>
>
>
--
Aldo Ugolotti, Ph.D.
Post-doc fellow
Materials Science Dept. U5,
Università degli Studi di Milano-Bicocca
via Cozzi 55,
20125 Milano (MI)
Italy
e-mail: a.ugolotti at campus.unimib.it
More information about the users
mailing list