Modeling a via profile etched in a CHF3/CF4 plasma using a neural network

Via profiles of oxide films were qualitatively modeled using a neural network. The oxide films were magnetically etched in a CHF3/CF4 plasma with various radio-frequency (RF) powers, pressures, and CHF3 and CF4 flow rates. A statistical 2(4-1) fractional factorial experiment was conducted to characterize the behavior of the via profile. The neural network was trained on nine experiments, and the trained model was evaluated on another eight experiments, not belonging to the training data. As a function of the training factors, the prediction accuracy of profile model was optimized, and the optimized model had a prediction error of 3.05degrees. Compared to the statistical regression model, this was about a 43% improvement in the prediction accuracy. Using the model, we made several 3-D) plots to unveil underlying etch mechanisms, including the factor interaction effects, involved in via formation. As expected, the profile angle decreased with increasing RF power without regard to the pressure. The DC bias induced by the pressure played an important role in affecting the profile angle. The profile became more positively sloped with increasing the CHF3 flow rate, contrary to what was noticed with the variation in the CF4 flow rate. For the profiles to be positively sloped, the effects of either pressure or CHF3 flow rate must be more noticeable than they are for the profiles to be negatively sloped.