Qualitative Interpretation of Etch Profile Nonuniformity Using a Wavelet and a Neural Network

A method to predict plasma etch profile nonuniformity is presented. This was accomplished by using a neural network and a wavelet. The wavelet was used to define a metric of the profile nonuniformity. The method was applied to the etching of tungsten films in a helicon SF₆ plasma. The etch process was characterized by a 2^4-1 fractional factorial experiment. The process parameters that were varied in the design include the radio-frequency source power, the bias power, the substrate temperature, and the SF₆ flow rate. The fluorine concentration [F] measured using optical emission spectroscopy was related to the profile nonuniformity. The model prediction accuracy was optimized as a function of training factors, and the optimized model had a root-mean squared error of 6.43 %. Using the optimized model, we qualitatively estimated etch mechanisms. Decreasing each process parameter generally reduced the profile nonuniformity. For variations either in the source power or temperature, both the profile nonuniformity and [F] were highly correlated. The presented method can be applied to characterize any plasma-processed surfaces.