That's a tricky question. On the one hand you want to take e.g. surface reconstruction into account, but on the other hand the TB model may not have this kind of flexibility. Some TB models have matrix elements that are functions of the distance between atoms and can at least attempt to take this into account, like Huckel and DFTB and some nearest-neighbor models. But the equilibrium of the model is usually chosen at the experimental lattice constant, so at least one should use that for the basic crystals, and only let surfaces relax. But DFTB on the other hand has quite good relaxation capabilities, so one needs to distinguish different TB models from each other.
