reliability of pseudopotentials for TM



 Concerning the accuracy of pseudopotentials for transition metals (TM),
 I'd like to make the following comments on the contribution of Jan
 Hrusak, Nov. 4, 92. In my opinion, his comments, especially the table
 summarizing some results for the iron metal, give a completely wrong
 picture concerning the reliability of the Hay-Wadt pseudopotentials.
 Using pseudopotentials one should keep in mind the following:
 1. In order to judge the accuracy of a pseudopotential one should
    compare results of pseudopotential to all-electron calculations
    performed at a corresponding level of theory, e.g. the HF level.
    One should always exclude basis set effects since pseudopotentials
    and basis sets are different things. I suggest to do such
    comparisions always at the finite-difference level. In case of
    the iron atom one can take the numerical all-electron data
    (nonrelativistic and spin-orbit averaged quasirelativistic)
    from Martin and Hay, JCP 75, 4539 (1981). Their Cowan-Griffin
    results for ionisation and excitation energies agree with values
    calculated by us using the related Wood-Boring scheme within 0.01 eV.
    It is not correct to compare UHF results (which give no well-defined
    LS-state anyhow) to such RHF results.
 2. Errors due to the valence basis sets can be checked by comparision
    of the corresponding results to finite difference pseudopotential
    results or at least pseudopotential calculations performed with
    very large basis sets (we used 15s15p15d before we applied the
    finite-difference code).
 3. Correlation errors can finally be checked by comparing correlated
    quasirelativistic pseudopotential calculations to experimental data.
    One should not forget to average the experimental numbers over
    spin-orbit components when doing the pseudopotential calculation
    in LS-coupling (Gaussian). This was not done by Jan Hrusak.
    There is much literature on correlation effects in transition metals
    and MP2 is definitely not the method to be used. I suppose that the
    excellent results Jan Hrusak obtains with his Ar-core pseudopotential,
    a 6s3p8d/5s3p5d basis (how can one describe 4p or 4s2->4p2 correlation
    with this basis set?) in the framework of MP2 are not due to an
    excellent pseudopotential, but rather to an excellent error
    cancelation (pseudopotential error, basis set error, correlation
    error, error in averaging the experimental numbers).
 I really have no idea how
 Jan Hrusak can 'publish' an all-electron excitation energy for iron
 for 5D->3F (most likely s2d6 -> d8) of 7.53 eV, compare
 this to pseudopotential results for 5D->3F (most likely s2d6 -> s1d7)
 of 2.08 eV (error 5.45 eV) and 3.37 eV (error 4.16 eV) and state
 (without any doubts, that his comparison is affected by errors), that
 this is due to the 'ground state oriented parametrization procedure
 more than with the valence space'. In order to discuss this, one
 needs to know frozen-core results for Ne and Ar core: The frozen-core
 errors in excitation and ionisation energies for 13 LS states of Fe and
 Fe(+) are less than 0.02 eV for the Ne core and less than 0.25 eV for
 the Ar core in finite difference all-electron HF calculations, when
 this core is taken from the s2d6 5D ground state. I don't see, how
 4 or 5 eV errors should arise from this in pseudopotential work.
 Therefore, in order to prove that Hay-Wadt pseudopotentials are quite
 accurate, in the following table I summarize some numbers for iron
 pseudopotentials of Hay and Wadt and ourselves (Dolg, Wedig, Stoll,
 Preuss, JCP 86, 866, 1987) in comparison to the all-electron data
 published by Martin and Hay:
                           5D -> 5F    5D -> 3F    5D -> 6D    5D ->
 4F
                         s2d6  s1d7   s2d6   d8  s2d6  s1d5  s2d6    d7
 num. RHF, nonrelativistic
 AE,HF (Martin,Hay)            1.80        7.46        6.28        7.95
 ECP (Ar, Hay,Wadt)            1.53        7.28        6.06        7.73
 ECP (Ne, Hay,Wadt)            1.96        7.72        6.28        8.12
 PP  (Ne, Dolg,Stoll,Preuss)   1.78        7.50        6.28        7.95
 with 7s6p5d/5s4p3d basis set  1.79        7.48        6.33        7.99
 num. RHF, quasirelativistic
 AE,HF (Martin,Hay)            2.06        7.87        6.34        8.33
 PP  (Ne, Dolg,Stoll,Preuss)   2.06        7.94        6.35        8.36
 with 7s6p5d/5s4p3d basis set  2.07        7.93        6.40        8.40
 (This is the data for 4 out of 13 states I looked on in 1986; the
 results for the other states are of the same quality.)
 It is clearly seen that all pseudopotentials give reliable results at the
 HF or SCF level. The same is true for the pseudopotentials published
 by the Christiansen group or the Toulouse group (I don't have the
 numbers for Fe, but e.g. for Ni the Christiansen pseudopotential HF
 results agree with all-electron HF results within 0.1 eV).
 Finally, I would like to remark that it is rather a matter of a proper
 initial guess and not a problem of the Hay-Wadt pseudopotential, if
 Gaussian calculations often converge to an excited state.
 The message should be:
 Distinguish pseudopotential, basis set and computational method
 as well as the errors occuring at in all of these. Although the
 final results might be 'excellent', it could be accidential.
 Michael Dolg
 Institut fuer Theoretische Chemie
 Universit{t Stuttgart
 Pfaffenwaldring 55
 W 7000 Stuttgart 80
 CFAT .at. DS0RUS1I