Use of a neural network to characterize the surface roughness of a SiC film

The surface roughness due to plasma etching was characterized using a neural network and atomic force microscopy. The proposed method was used to etch silicon carbide for fabricating a high-power device. The etching process was systematically characterized by means of a 2⁵ full factorial experiment. With the use of a genetic algorithm, the prediction performance of the neural network was optimized and the resulting prediction error was 0.108 nm. The effects of the plasma parameters were examined by predicting and experimentally validating the surface roughness under various plasma conditions. The surface roughness strongly depended on the bias power. For variations in the source power, it was highly correlated with the dc bias. As the wafer was placed closer to the plasma source, the surface roughness came to be dominated by chemical reactions rather than by ion bombardment. The proposed technique can be applied to constructing a prediction model for any complex plasma process.