Prediction of plasma-induced DC bias using polynomial neural network

DC bias plays an important role in characterizing or controlling plasma processes. A predictive DC bias model is constructed using a polynomial neural network (PNN) and a genetic algorithm (GA). The GA was used to optimize PNN training factors, including the number of input variables to a partial description (PD), the selection of input variables, and the type of polynomial for PD. The DC bias data were collected during the etching of silicon carbide in a C2F6 inductively coupled plasma. The process parameters involved are a radio frequency (rf) source power, bias power, pressure, gap, and O2 fraction. The etch process was characterized by a 25 full factorial experiment. Additional 17 experiments were conducted to test the predictive performance of constructed models. Compared to statistical regression models, GA-PNN demonstrated a drastic improvement in predicting DC bias under various plasma conditions. The GA-PNN can generally be applied to model other plasma processes.