Process monitoring and feedback control using multiresolution analysis and machine learning

Online process monitoring and feedback control are two widely researched aspects that can impact the performance of a myriad of process applications. Semiconductor manufacturing is one such application that due to the ever increasing demands placed on its quality and speed holds tremendous potentials for further research and development in the areas of monitoring and control. One of the key areas of semiconductor manufacturing that has received significant attention among researchers and practitioners in recent years is the online sensor based monitoring and feedback control of its nanoscale wafer fabrication process. Monitoring and feedback control strategies of nanomanufacturing processes often require a combination of monitoring using nonstationary and multiscale signals, and a robust feedback control using complex process models. It is also essential for the monitoring and feedback control strategies to possess stringent properties such as high speed of execution, lowcost of operation, ease of implementation, high accuracy, and capability for online implementation. Due to the above requirement, a need is being felt to develop state-of-the-art sensor data processing algorithms that can perform far superior to those that are currently available both in the literature and commercially in the form of softwares. The contributions of this dissertation are three fold. It first focuses on the development of an efficient online scheme for process monitoring. The scheme combines the potentials of wavelet based multiresolution analysis and sequential probability ratio test to develop a very sensitive strategy to detect changes in nonstationary signals. Secondly, the dissertation presents a novel online feedback control scheme. The control problem is cast in the framework of probabilistic dynamic decision making, and the control scheme is built on the mathematical foundations of wavelet based multiresolution analysis, dynamic programming, and machine learning.Analysis of convergence of the control scheme is also presented. Finally, the monitoring and the control schemes are tested on a nanoscale manufacturing process (chemical mechanical planarization, CMP) used in silicon wafer fabrication. The results obtained from experimental data clearly indicate that the approaches developed outperform the existing approaches. The novelty of the research in this dissertation stems from the fact that they further the science of sensor based process monitoring and control by uniting sophisticated concepts from signal processing, statistics, stochastic processes, and artificial intelligence, and yet remain versatile to many real world process applications