Fundamental issues in iterative learning controller design: Convergence, robustness, and steady state performance

Development of photolithography technology is pivotal for the next generation of semiconductor manufacturing. Hence, tighter process bounds are necessary to guarantee throughput and yield requirements, which necessitates the design of motion control systems with faster and more accurate positioning capability. Since the process of photolithography is repetitive, control techniques that exploit the repetitiveness to enhance performance should be developed. This dissertation focuses on design of Iterative Learning Controllers (ILC) for wafer scanning systems.;ILC is a control technique that iteratively fine-tunes the feedforward signal by considering the error from previous runs of the repetitive process. For the same feedback system, different configurations of ILC are possible depending on the choice of the signal that is used to learn and the point of injection of the learned signal. While these alternative structures are mathematically equivalent, some structures provide better numerical stability because the transfer functions involved in design of the ILC system avoid inifinite or null gains at any frequency.;Typically, optimization problems are solved by iterative methods. Similarly, an ILC algorithm is designed for iterative minimization of a cost function (tracking error). The key difference is that the descent direction is determined by an experiment run. Therefore, there exist very interesting parallels between ILC and optimization algorithms. We extend this analogy to ILC with iteration-varying cost function optimization. This is inspired humans learning patterns: rapid learning in early stages and then slower learning with experience. As a result of this dynamic rate of learning, faster and more robust learning is possible.;Finally, a framework for design of ILC algorithms for repetitive processes with nonrepetitive events is proposed. A stochastic analysis of the ILC system with nonrepeating events shows that there is a tradeoff between the rate of learning and the steady state error. We propose ILC algorithms that use the structure of the repetitive disturbances to improve robustness to nonrepetitive events. Further, repeatable components in the error decay exponentially over iterations, while nonrepeatable components do not. This leads us to hypothesize that the rate of learning should be slowed after many iterations as nonrepeating components start dominating over repeating ones.