The program starts with a calculation of the free atoms, assuming spherical symmetry. The formation energy is calculated w.r.t such atoms. You have to specify the configuration (i.e. which orbitals are occupied) in the Dirac subkey of the block key AtomType, and you can for instance use the experimental configuration. Keep in mind, however, that this is not necessarily the optimal configuration for your density functional. For instance, Ni has experimentally two electrons in the 4s shell, but with LDA you will find that it is energetically more profitable to move one electron from the 4s to the 3d. The configuration of the reference atoms does not (i.e. should not) affect the final (SCF) density.
Besides the available basis sets in $ADFHOME/atomicdata/Band, you could in principle use the basis functions from the database of the molecular ADF program (see the documentation of ADF for how this database is organized). The functions you will find there are STOs, which is not optimal since BAND offers you the option to use NOs from the numerical atom. The most efficient approach is to use the NOs and remove from the ADF basis set those STOs that are already well described by the NOs.
As an example we will construct a basis for the Ni atom with orbitals frozen up to the 2p shell, derived from a triple-zeta ADF basis. In the Dirac subkey of the block key AtomType you specify that the NOs up to 2p should be kept frozen and that the 3d and 4s NOs be included in the valence basis. Copy from the ADF database all 3d, 4s and the polarization functions into the BasisFunctions subkey of the block key AtomType and remove the middle STOs of the 3d and the 4s.
Usually it is already quite adequate for a good-quality basis to augment each NO with one STO. You could then take a double zeta ADF basis and remove one of the 3d and one of the 4s STOs. We often find that such a basis, with one STO added per NO, has a quality that is comparable to triple zeta STO sets. We strongly recommend that you use combined NO/STO bases. Of course, you may want to verify the quality of the basis set by calculations on a few simple systems.
You can copy the fit functions from the ADF database into the FitFunctions subkey of the block key AtomType. As a matter of experience (and justified by a somewhat different handling of fit functions between the two programs), BAND is in most cases (we have not yet seen an exception) less sensitive to the quality of the fit set than ADF is.
