Use of neural network to model the deposition rate of PECVD-silicon nitride films

Using a neural network, the deposition rate of silicon nitride films was modelled. The deposition process with a plasma-enhanced chemical vapour deposition system was characterized by a face-centred central composite circumscribed design. Six process parameters were involved, including the radio frequency power, pressure, substrate temperature and three gases (SiH4, NH3 and N-2). The prediction performance of the generalized regression neural network was significantly improved by optimizing the multi-parametrized training factors using a genetic algorithm. The optimized model had a root-mean-squared prediction error of 8.26 angstrom min(-1), which is an improvement of about 66% over the conventional model. From the optimized model, three-dimensional plots were generated to qualitatively interpret parameter effects. The predicted variations were partly validated with experimental data. For the variations in the temperature, the deposition rate was strongly correlated to the hydrogen concentration. The physical ion bombardment effect was pronounced at smaller hydrogen (H) concentrations. A comparison with the refractive index model facilitated identifying empirical relationships between the deposition rate and the formation of a Si- or N-rich film.