Scftargets

Target values for criteria for determining whether the SCF has converged, whose values are stored in a Numeric array of dimension n x m, where n is the number of geometry optimisation cycles (1 for a single point calculation) and m is the number of criteria.

The criteria vary from program to program, and depending on the program they may be constant for the whole of a geometry optimisation or they may be different for each geometry optimisation cycle. A more complete description follows.

Gaussian

 * The RMS change in the density matrix elements. Default is 1.0E-4 (1.0E-8 for geo opts).
 * The maximum change in the density matrix elements. Default is 1.0E-2 (1.0E-6 for geo opts).
 * The change in energy. Default is 5.0E-05 (1.0E-06 for geo opts).

GAMESS

 * Maximum and root-mean-square (RMS) density matrix change: starts from 5.0E-05 by default and changes over the course of a geometry optimisation. ROHF calcualtions use SQCDF instead of the standard RMS change.

Jaguar 4.2
The Jaguar 4.2 targets depend on whether it is a geometry optimisation or not:
 * For geometry optimisation and hyper/polarisability, the rms change in the density matrix elements is used as a criterion (controlled by the dconv keyword). Default is 5.0E6.
 * The energy convergence criterion (keyword econv) is ignored for geometry optimisation calculations but is used for SCF calculations. Default is 5.0E5, except for hyper/polarisability where it is 1.0E6.

(Note sure how correct this is - but it's taken from the manual)

ADF

 * Maximum element A < target criterion
 * Norm A < 10 * target criterion
 * where A is the commutator of the Fock and P matrices

The value of the target criterion depends on the calculation:


 * The "hard" target, SCFconv, is used for single point calculations, and the last step of geometry optimizations. It defaults to 1.0E-6, and can be changed in the SCF keyword block.


 * The "soft" target, scfconv2, is used for the first step of a geometry opimization. It defaults to 1.0E-3, and can also be changed in the SCF keyword block.


 * For intermediate steps in a geometry optimization the situation is more complicated and depends on the gradient and the integration accuracy. After a post on the ADF user's forum, I found that it is calculated as follows:


 * new_convcriteria = max(SCFconv, min(old_convcriteria, grdmax/30, 10**(-accint))) )


 * where old_convcriteria is the value from the previous geometry cycle, grdmax is the max gradient from the last geometry cycle and accint is the current integration accuracy.

(There is some information in the ADF manual here.)

GAMESS-UK
According to http://www.cfs.dl.ac.uk/docs/gamess_manual/chap4/node6.html, SCF convergence is determined by convergence of the elements of density matrix. The default value is 1E-5, but it appears to be 1E-7 for geoopts.