[CPMD-list] Re: free isolated atom energiesB (fwd)
Axel Kohlmeyer
axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Sat Sep 18 20:53:56 CEST 2004
>>> "EM" == Eduardo Ariel Menendez P <Eduardo> writes:
EM> Hello again,
hello eduardo,
EM> First of all, thanks to AK, AS, and OY for your comments. Then a few more
EM> calculations and questions.
thank you for raising this issue. i think this is a useful discussion
for many people in order to learn what you can do with plane-wave
pseudopotential based DFT methods (and what not).
one should keep in mind, though, that the CPMD code is especially
well suited to do Car-Parrinell molecular dynamics, where single
point calculation accuracy issues at least partially cancel out.
yet one has to see, that many of the parameters used in the past
(and thus the default settings in CPMD) were chosen to get the
simulations done at all. it is only due to the ever increasing
computational power, that one can affort to run more 'expensive'
calculations and there are already several publications
(see. e.g. j.chem.phys.B 108(2004), pp12990-12998) that re-evaluate
those parameters with respect to accuracy.
>>
>> stupid question: did you already try the equivalent pseudopotentials
>> from the goedecker hartwigsen library? at least for LDA there is a very
>> large collection, including Te and two version of Cd potentials (normal
>> and semi-core). also the way they are defined seems to make them much
>> more suitable for those elements than the troullier-martins
>> pseudopotentials in separable form. at least this is my (limited)
>> experience so far.
EM> I have used them (Goedecker's PPs), but I find that the cutoff required
EM> for convergence are larger than for MT pseudos. For O I find the same
EM> behavior in the
EM> atomic calculation. However, see below.
i agree. the high cutoff can be very painful, and for oxygen a
troullier-martins pseudopotential should perform as well requiring
a lower cutoff. my remark was with respect to the fact, that you
cannot have multiple projectors with the same angular momentum
in the TM/KB-scheme.
since you care about accuracy and for the elements you mentioned,
the goedecker approach should give a more accurate description
(as does the vanderbilt approach, btw). at least, this is my
personal experience.
>>
>>
>> USPPs and LANCZOS are not compatible. see the recent
>> mail by juerg hutter in the mailing list archives.
EM> Let's turn to Norm conserving pseudopotentials.
>>
>> finally, i am not so sure, if you can compare the
>> absolute values, since you are using different grids
>> for the atomic and the cpmd calculation.
EM> Naive thought: I guess I should obtain comparable results, if one wishes
EM> to trust in CPMD
EM> molecular dynamics. ?What about the forces? At least, the Born-Oppenheimer
EM> MD is likely to give wrong forces?
well, with respect to forces but also the eigenvalues you should
converge your wavefunction more tightly. the default for CONVERGENCE
ORBITALS is 1.0e-5 which is good enough to start a CP-MD simulation
(the coefficients are oszillating around the optimum anyways
and if you thermostat the electronic degrees of freedom you will
basically reoptimize your wavefunction implicitely during the simulation).
but to get converged forces (and KS energies) you need to converge
as far as (numerically) possible, 1.0e-7 or lower if possible.
if you don't converge that well, you will have a significant
energy drift in BO-MD simulations.
EM> Ari Seitsonen pointed that
EM> " 1) Like you noticed, the occupation numbers 2, 2, 1, 1 do not make really
EM> sense; well, this is leads to a non-spherical atom (p_x, p_y and p_z
EM> are occupied differently), which is used in the pseudo potential
EM> program. You should use an equal occupation of the p levels if you want
EM> to compare with Paolo Giannozzi's program (hmm, keyword 'OCCUPATION'?)"
EM> Then I procceeded to calculate the O atom (with Goedecker's PP) keeping
EM> the 3 2p orbitals with
EM> a population of 4/3. I found convergence with the following sequence:
EM> 1)
EM> &CPMD
EM> OPTIMIZE WAVEFUNCTION
EM> MAXSTEP
EM> 150
EM> CENTER MOLECULE ON
EM> &END
EM> &DFT
EM> NEWCODE
EM> FUNCTIONAL LDA
EM> LDA CORRELATION PZ
EM> SLATER
EM> 0.666667
EM> GC-CUTOFF
EM> 0.1E-07
EM> &END
EM> &SYSTEM
EM> SYMMETRY
EM> 0
EM> POISSON SOLVER TUCKERMAN
EM> ANGSTROM
EM> CELL
EM> 10.0 1.0 1.0 0 0 0
EM> CUTOFF SPHERICAL
EM> 155.0
EM> STATES
EM> 4
EM> OCCUPATION FIXED
EM> 2.0 1.33333333333333 1.33333333333333 1.33333333333333
EM> &END
EM> &ATOMS
EM> *O-q6
EM> LMAX=P LOC=P
EM> 1
EM> 0.000000 0.000000 0.000000
EM> &END
hmmm. this input should not work. my version stops and the manual states,
OCCUPATION FIXED only works in combination with a diagonalization
scheme. you can basically do everything in one sweep by using
LANCZOS, NSTATES 6, and OCCUPATION FIXED directly for the
wavefunction optimization (takes a little longer, but that
does not matter here). since your cell is large and you don't
have (or want) a charge or dipole, the poisson solver can be
omitted as well.
[...]
EM> Then, I got the eigenvalues:
EM> EIGENVALUES(EV) AND OCCUPATION:
EM> 1 -23.6274394 2.000 2 -9.0449296 1.333
EM> 3 -9.0448316 1.333 4 -9.0448264 1.333
EM> 5 -0.2474729NC 0.000 6 1.2166765NC 0.000
EM> CHEMICAL POTENTIAL = 1.2166765461 EV
EM> Eigenvalues 5 and 6 are not converged yet, but at least there are 3 almost
EM> equal energies close to the reference 2p energy, as well as the
AM> 2s -0.871362 Ha= -23.70104 eV
AM> 2p -0.338381 Ha = -9.20 eV
if i converge the wavefunction up to 5.0e-8 i get:
EIGENVALUES(EV) AND OCCUPATION:
1 -23.6689855 2.000 2 -9.0841701 1.333
3 -9.0841701 1.333 4 -9.0841701 1.333
5 -0.2755064NC 0.000 6 1.2664967NC 0.000
CHEMICAL POTENTIAL = 1.2664966540 EV
adding 200 more iterations with a KS-energies run gives:
EIGENVALUES(EV) AND OCCUPATION:
1 -23.6689855 2.000 2 -9.0841701 1.333
3 -9.0841701 1.333 4 -9.0841701 1.333
5 -0.2764696 0.000 6 1.2588665NC 0.000
CHEMICAL POTENTIAL = 1.2588664817 EV
cranking up the cutoff to 200ry yields:
EIGENVALUES(EV) AND OCCUPATION:
1 -23.6516718 2.000 2 -9.0905575 1.333
3 -9.0905574 1.333 4 -9.0905574 1.333
5 -0.2722334 0.000 6 1.2700999NC 0.000
EM> There are still differences with the reference energies that are larger
EM> than the standard criterium of 0.01 ev/atom.
EM> Are this errors not important for the calculation of forces?
the eigenvalues are specifically important if you want to check
your pseudopotential for transferability (e.g. if you compare with
results for 0+ or O2+). to test the the resulting forces, you can
do geometry optimizations of small molecules or calculations of
lattice constants etc.
EM> Is there any other simple test for the new generated pseudopotentials
EM> after being translated to the CPMD format?
well, the most important tests, you can do even before doing
cpmd calculations. this is the tests for ghost states and
transferability. the fhi98PP package has a very nice description
of what can and has to be done.
best regards,
axel.
EM> Best regards,
EM> Eduardo
--
=======================================================================
Axel Kohlmeyer e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53 Fax: ++49 (0)234/32-14045
D-44780 Bochum http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.
More information about the CPMD-list
mailing list