Contemporary orchestration practice harbours a number of aesthetic inquiries relating to the employment and arrangement of percussion instruments. Due in part to the fact that percussion instruments largely occupy an inharmonic timbre space, they encompass a diverse and distinctly nuanced musical idiom in comparison to harmonic instruments, both in terms of their textural interplay, musical function and cultural significance. In response to this perspective, we present a neural network approach to the parameter estimation of physically modelled, abstract percussion instruments. The approach presented here serves as our initial attempt towards creating a computer-assisted orchestration methodology specifically targeting the musical employment and arrangement of inharmonic timbres and percussive instruments. The neural architecture presented here has been trained and tested using a pair of two-dimensional physical models, to gauge a sense of our architecture's successes and limitations as we continue to expand this approach to include more two-dimensional models. This works poses as our first technological inquiry into this field, which has here been quantitatively assessed, with plans to undertake more rigorous and comparative tests in the near future.