A neural-network-based approach to determining a robust process recipe for the plasma-enhanced deposition of silicon nitride thin films

We consider the problem of locating a process recipe that produces outputs which are, in some sense, least sensitive to small fluctuations in the process condition. Specifically, we determine the most robust process recipe for the plasma-enhanced chemical vapor deposition of silicon nitride thin films having specified optical properties. An appropriate sensitivity functional describing the relationship between the process inputs and outputs and their localized variability is defined in terms of response surfaces and the response surface gradients. Determining the most robust process recipes which produce films with a given refractive index is then formulated as a constrained minimization problem. The silicon nitride films are characterized by spectroscopic ellipsometry and the requisite response surfaces are obtained by training feedforward artificial neural networks with available data. Numerical findings are presented, validated via simulation, and discussed.