A neural network is developed to fit a potential energy surface of silicon derived from Frauenheim tight-binding data for silicon. The tight-binding method retains the essentials of quantum mechanics for electronic structure calculations but is faster to calculate than a full ab initio model. The development of the neural network potential energy surface was carried out by a progressive refinement of the design parameters. The refinement of the models went hand in hand with the difficulty encountered in developing a transferable network potential. Both equilibrium and non-equilibrium parts of the potential energy surface were represented in the training data set. The neural network potential was fitted on dimers, linear and angled trimers, tetramers, diamond structures, distorted diamond lattice systems, and the BC8, ST12, BCT5 and β-tin structures.