There is no way in general from ATK or any DFT code to just push a button and you get a tight-binding model. Creating a new TB model requires a fitting procedure where the DFT results (or some other data, could be experimental) are used as target values and you adjust the parameters until you have as close match as possible. This is a high-dimensional, highly non-linear problem with lots of local minima which requires a careful approach and the use of special techniques like genetic algorithms.
An interesting, slightly different approach can be found in PRB 84, 155119 (2011), or the papers on DFTB by in particular Frauenheimer.
