Iterative Learning Control: Feasibility and Implementation for a Reticle Stage

ASML is always searching for improvements in reducing the servo error of both the Wafer Stage and the Reticle Stage of their lithography machines. Due to the repetitive motion and corresponding repetitive servo error of the Reticle Stage Short Stroke (RSSS), Iterative Learning Control (ILC) can be a powerful method to further reduce the servo error. ILC is an iterative optimization procedure, where the servo error data from the previous trial is used to create an additional feedforward input vector to further reduce the servo error. The main goal of this MSc thesis project is the implementation of ILC on the RSSS of the ASML Twinscan NXT wafer scanner, with the purpose of improving the Moving Average (MA) and Moving Standard Deviation (MSD) of the servo error during the scanning window. Successful application of common ILC methods requires exact repetitiveness of the setpoint trajectory as well as the need to iterate. However, the amount of iterations needed for the ILC must be kept at a minimum to maintain a profitable throughput of wafers. Furthermore, the setpoint trajectories of the machine can be significantly different for different jobs. Frequency-based ILC, which is the most straightforward method of designing the ILC algorithm, is able to improve the MA and MSD of the servo error in the y-direction (i.e., scanning direction) of the RSSS with 81% and 21%, respectively. In the x-direction, improvements in the MA and MSD of respectively 50% and 14% are made. In both directions the performance rise is achieved after just one iteration. Furthermore, an improvement by ILC in one direction does not cause any significant performance decrease in another direction which will drive them out of specifications. Frequency-based ILC is, however, not able to cope with changes of the setpoint during scanning, yielding significant performance loss. Norm-Optimal ILC with the use of basis functions is an ILC method where the generated feedforward signal is parameterized in terms of the setpoint. If the changes in the setpoint are known the ILC algorithm can adapt to this change and thus keep its performance. Norm-Optimal ILC with basis functions outperforms standard ILC algorithms by roughly tenfold when a large setpoint change is performed during operation of the RSSS. For an improvement in the MSD of the servo error, extra sinusoidal functions with specific frequencies are added to the basis which will prevent amplification of the servo error around those frequencies in the MSD domain. Norm-Optimal ILC with basis functions has proven its robustness against ?setpoint changes when applied on the RSSS, but the trade-off against general performance is still too big. If the general performance can be improved further, especially in the MA region of the servo error, Norm-Optimal ILC with basis functions can be a valuable addition to the RSSS control structure when dealing with changing setpoint profiles. Iterative Learning Control proves to give significant and valuable improvement of the performance of the RSSS in terms of overlay (MA) and fading (MSD) and can be implemented on top of the current control architecture at ASML.Systems and ControlDelft Center for Systems and ControlMechanical, Maritime and Materials Engineerin