4.5 Converging SCF Calculations

4.5.10 Small-Gap Systems

SCF calculations for systems with zero or small HOMO-LUMO gap (such as metals) can exhibit very slow convergence or may even fail to converge. This problem arises because the energetic ordering of orbitals and states can switch during the SCF optimization leading to discontinuities in the optimization. Using fractional MO occupation numbers can improve the convergence for small-gap systems. In this approach, the occupation numbers of MOs around the Fermi level are allowed to assume non-integer values. This “occupation smearing” allows one to include multiple electron configurations in the same optimization, which improves the stability of the optimization.

We follow the pseudo-Fractional Occupation Number (pFON) method of Rabuck and Scuseria763 that scales the MO occupation used to construct the AO density:

Pμν=p=1NnpCμpCνp. (4.42)

For a conventional (integer occupation number) SCF run, the occupation number np is either one (occupied) or zero (virtual). In pFON, the occupation numbers are following a Fermi-Dirac distribution,

np=(1+e(ϵp-ϵF)/kT)-1, (4.43)

where ϵp is the respective orbital energy and kT the Boltzmann constant and temperature, respectively. The Fermi energy ϵF is set to (ϵHOMO+ϵLUMO)/2 in our implementation. To ensure conservation of the total number of electrons, the pFON approach re-scales the occupation numbers so that pnp=Nel.

There are several parameters to control the electronic temperature T throughout a pFON SCF run. The temperature can either be held constant at finite temperature (Tinit = Tfinal), or the system can be cooled from a higher temperature down to the final temperature. So far, no zero-temperature extrapolation has been implemented.

OCCUPATIONS
       Activates pFON calculation.
TYPE:
       INTEGER
DEFAULT:
       0
OPTIONS:
       0 Integer occupation numbers 1 Not yet implemented 2 Pseudo-fractional occupation numbers (pFON)
RECOMMENDATION:
       Use pFON to improve convergence for small-gap systems.

FON_T_START
       Initial electronic temperature (in K) for FON calculation.
TYPE:
       INTEGER
DEFAULT:
       1000
OPTIONS:
       Any desired initial temperature.
RECOMMENDATION:
       Pick the temperature to either reproduce experimental conditions (e.g. room temperature) or as low as possible to approach zero-temperature.

FON_T_END
       Final electronic temperature for FON calculation.
TYPE:
       INTEGER
DEFAULT:
       0
OPTIONS:
       Any desired final temperature.
RECOMMENDATION:
       Pick the temperature to either reproduce experimental conditions (e.g. room temperature) or as low as possible to approach zero-temperature.

FON_NORB
       Number of orbitals above and below the Fermi level that are allowed to have fractional occupancies.
TYPE:
       INTEGER
DEFAULT:
       4
OPTIONS:
       n number of active orbitals
RECOMMENDATION:
       The number of valence orbitals is a reasonable choice.

FON_T_SCALE
       Determines the step size for the cooling.
TYPE:
       INTEGER
DEFAULT:
       90
OPTIONS:
       n temperature is scaled by 0.01n in each cycle (cooling method 1) n temperature is decreased by n K in each cycle (cooling method 2)
RECOMMENDATION:
       The cooling rate should be neither too slow nor too fast. Too slow may lead to final energies that are at undesirably high temperatures. Too fast may lead to convergence issues. Reasonable choices for methods 1 and 2 are 98 and 50, respectively. When in doubt, use constant temperature.

FON_E_THRESH
       DIIS error below which occupations will be kept constant.
TYPE:
       INTEGER
DEFAULT:
       4
OPTIONS:
       n freeze occupations below DIIS error of 10-n
RECOMMENDATION:
       This should be one or two numbers bigger than the desired SCF convergence threshold.

FON_T_METHOD
       Selects cooling algorithm.
TYPE:
       INTEGER
DEFAULT:
       1
OPTIONS:
       1 temperature is scaled by a factor in each cycle 2 temperature is decreased by a constant number in each cycle
RECOMMENDATION:
       We have made slightly better experience with a constant cooling rate. However, choose constant temperature when in doubt.

Example 4.13  pFON calculation of a metal cluster.

$molecule
   0 1
   Pt        -0.20408        1.19210        0.54029
   Pt         2.61132        1.04687        0.66196
   Pt         0.83227        0.03296       -1.49084
   Pt         0.95832       -1.05360        0.92253
   Pt        -1.66760       -1.07875       -1.02416
$end

$rem
   METHOD           pbe
   ECP              fit-lanl2dz
   SYMMETRY         false
   OCCUPATIONS      2   ! pseudo-fractional occupation numbers
   FON_NORB         10
   FON_T_START      300 ! electronic temperature: 300 K
   FON_T_END        300
   FON_E_THRESH     5   ! freeze occupation numbers once DIIS error is 10^-5
   GEN_SCFMAN       false
$end