Modeling of Silicon Carbide ECR Etching by Feed-Forward Neural Network and Its Physical Interpretations

The behavior of electron cyclotron resonance etching of 4H-SiC based on SF₆ + O₂ plasma is studied using a three-layered feed-forward neural network model trained by the Broyden, Fletcher, Goldfarb, and Shanno optimization algorithm. The etch rate of 4H-SiC is modeled as a function of microwave power, dc bias, process pressure, and O₂/(SF₆ + O₂) flow ratio referred to as the O₂ fraction. The results clearly reveal the interaction of the various process parameters and their resultant effects on the etch rate. It is found that, by varying the O₂ fraction and process pressure, optimized etch rate peaks can be achieved. Increasing dc bias and microwave power is found to result in the optimized etch rate peaks occurring at higher O₂ fraction and process pressure, respectively. In addition, increasing dc bias or microwave power will increase the etch rate. However, there is saturation in the etch rate at higher dc bias, which does not occur with microwave power. Based on these modeling results, detailed physical interpretations of the etching process are given in terms of the chemical reaction of 4H-SiC with F, O, positive ion bombardment, etc.