The LINEARSCALING keyword has a very similar function to the TAILS keyword described above. In addition to defining the precision of operations related to operations in the numerical integration grid, it also defines the precision for the calculation of the overlap matrix, the fit integrals, and the density fit procedure. Default values have been chosen which result in negligible differences in the results for our test calculations, so that these defaults can be considered safe. They have been chosen similarly to the defaults for the TAILS keyword.
However, it may be advisable to modify the settings for the linear scaling parameters in two cases. First, if a very accurate result is needed, and numerical noise is to be completely eliminated, strict values can be specified. Especially for small molecules, where timings are not so large anyway, this may be of interest. Second, for large molecules, in which the calculations are very time-consuming, one can experiment with less strict values for the LINEARSCALING block keyword. In such a case one should be aware of the reduced accuracy and preferably test the influence of the changes on the results.
In the simplest application of the LINEARSCALING keyword, only one parameter is provided. All the subkeys described below will then be given this value. A very large value implies a calculation where no distance cut-offs are used. A normal value (almost default situation) would be 8 for linscal, 6 gives a faster but somewhat sloppier result. Whether this is acceptable is strongly case-dependent. A value of 10 or 12 is already quite strict and, unless there are some sort of numerical problems, there should not be much influence on the results by choosing a stricter value than that. A value of 99 for linscal virtually excludes the possibility that something will be neglected.
LINEARSCALING linscal
More refined control is possible by using the full block key
LINEARSCALING
CUTOFF_FIT epsfit
OVERLAP_INT ovint
PROGCONV progconv
CUTOFF_COULOMB epsvc
CUTOFF_MULTIPOLES epsmp
END
CUTOFF_FIT
determines how many atom pairs
are taken into account in the calculation of the fit integrals and the density
fit procedure. If the value is too low, charge will not be conserved and the
density fitting procedure will become unreliable. This parameter is relevant
for the timings of the FITINT and RHOFIH routines of ADF.
The default value for epsfit is accint + 4 (typically 8) and, where accint is the value specified for numerical
integration accuracy (see INTEGRATION keyword). This implies that the cut-off
criteria are automatically made more strict if a higher numerical integration
accuracy is chosen. The same is true for the other parameters in the
LINEARSCALING block.
OVERLAP_INT
determines the overlap criterion for pairs of AO's in the calculation of the Fock-matrix in a block of points. Indirectly it determines what the cut-off radii for AO's should be. The value of ovint has a strong influence on the timing for the evaluation of the Fock matrix, which is very important for the overall timings. The default value for ovint is accint + 2 (typically 6). Again, a higher value implies a safer but slower calculation.
PROGCONV
determines how the overall accuracy changes during the SCF procedure ('progressive convergence'). The idea is that one might get away with a lower accuracy during the initial SCF cycles, as long as the last cycle(s) is/are sufficiently accurate. The current default is that progconv has the value 0, which means that the accuracy in the beginning of the SCF is the same as in the rest of the SCF. This keyword is currently still in the testing phase, so we do not recommend changing its default value. The value of progconv determines how much lower the other parameters in the LINEARSCALING input block are at the beginning of the SCF than at the end.
CUTOFF_COULOMB
determines the radii for the fit functions in the evaluation of the (short-range part of) the Coulomb potential. As the Coulomb potential may take a sizable amount of time, the value chosen for epsvc may influence the total ADF timing signifcantly as well. The default value for epsvc is accint + 4 (typically 8).
CUTOFF_MULTIPOLES
determines the cut-offs in the multipole (long-range) part of the Coulomb potential. This term scales quadratically with system size, but has a small prefactor. In most cases, change in the epsmp value will not affect the CPU time significantly. The default value for epsmp is accint + 4 (typically 8).
