Process monitoring, modeling, and control of plasma etch systems

Selective etching of silicon dioxide over silicon is a frequently used process in the manufacture of semiconductor devices. Although central to the microelectronics manufacturing process, control strategies for plasma etch systems have been limited to statistical based process control and recipe control techniques, mainly due to a lack of in-situ real-time measurements of process performance. This dissertation focuses on the design, characterization, and implementation of two diagnostics on a research plasma reactor, and their use for process monitoring, empirical model building and advanced process control.;The diagnostics added to the reactor during this research included laser interferometry and voltage and current probes. An algorithm was developed to compute etch rate and end-point condition from the laser interferometer signal in near real-time. The RF monitoring sensor measured information about the RF voltage, current, and phase angle at three locations in the power delivery system--before and after the matching network and at the lower electrode. Transmission line analysis showed the importance of accurate characterization of stray capacitance and inductance in the power delivery system. Plasma parameters of impedance, delivered power, and sheath thickness were computed using simple equivalent circuit models for the plasma discharge. Measurement of fundamental and harmonic components of the RF voltage, current, and phase showed that the power generated in the plasma at harmonic frequencies was approximately 3% of the generator power.;These diagnostics provided the foundation for steady-state and dynamic model development of the plasma etch process. Several linear and nonlinear steady-state techniques including regression, neural networks, and projection of latent structures (PLS) were used in empirical model building. Dynamic models were also developed using neural network techniques. It was found that both the regression and recurrent neural network model structures provided a satisfactory fit of the data for the operating space investigated, although neural network models provided a dynamic model in a nonlinear state-space formulation. PLS techniques were used for model reduction and variable selection. The most relevant variables for model construction were power, pressure, and chamber impedance. The impedance measurement significantly improved the predictive capability of the model.