The use of bond centred fitting functions has a dual purpose. Firstly
it allows better modelling of the charge distribution in a bond
centre, which is normally where the valence charge is localised in a
cluster. Secondly it removes the need to include *d*- orbital fitting
functions on the bulk Si atoms. This is because the bond centred
Gaussians act on their nearby atomic sites with *s*, *p*, *d*,
etc. character, the strength of this effect tailing off with distance
between the Gaussian and the atomic site and increasing quantum
number. In order to maximise *d*-like character it is possible to
arrange these fitting functions symmetrically around an atomic site so
that their *p*-character exactly cancels; the minimum number of
functions required for this is an icosohedral arrangement. We tried
this for some of the oxygen diffusion calculations, where bonds are
breaking and reforming and so it is not always obvious where bond
centred functions should be located (see Chapter 6),
however in practise they did not increase the accuracy of the
calculation and significantly slowed calculation time.

However bond centred Gaussian functions bring associated disadvantages. Such additional fitting functions will locally improve the wavefunction and charge density modelling, thus lowering the cluster energy. This makes it harder to compare different structures, where the energy may change simply because the bond centred functions are no longer in such optimal locations. This has proved to be a particular problem in the case of oxygen in silicon where the cluster energy is critically dependant on the number and location of bond centred fitting functions. They often have little effect on final structure but the improved local charge description can improve the local vibrational modes.