Beyond Performance of Learning Control Subject to Uncertainties and Noise: A Frequency-Domain Approach Applied to Wafer Stages

• Published in: IEEE/CAA journal of automatica sinica Vol. 12; no. 1; pp. 198 - 214
• Main Authors: Song, Fazhi, Cui, Ning, Chen, Shuaiqi, Zhang, Kai, Liu, Yang, Chen, Xinkai, Tan, Jiubin
• Format: Journal Article
• Language: English
• Published: Piscataway Chinese Association of Automation (CAA) 01.01.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Adaptive learning; Convergence; Data models; Extended state observer; Feedforward systems; Frequency domain analysis; Integrated circuit modeling; Integrated circuits; Learning; learning control; model uncertainties; motion control; Noise; Noise control; Noise measurement; Semiconductor device modeling; State observers; stochastic noise; Stochastic processes; Uncertainty
• DOI: 10.1109/JAS.2024.124968
• ISSN: 2329-9266

The increasingly stringent performance requirement in integrated circuit manufacturing, characterized by smaller feature sizes and higher productivity, necessitates the wafer stage executing a extreme motion with the accuracy in terms of nanometers. This demanding requirement witnesses a widespread application of iterative learning control (ILC), given the repetitive nature of wafer scanning. ILC enables substantial performance improvement by using past measurement data in combination with the system model knowledge. However, challenges arise in cases where the data is contaminated by the stochastic noise, or when the system model exhibits significant uncertainties, constraining the achievable performance. In response to this issue, an extended state observer (ESO) based adaptive ILC approach is proposed in the frequency domain. Despite being model-based, it utilizes only a rough system model and then compensates for the resulting model uncertainties using an ESO, thereby achieving high robustness against uncertainties with minimal modeling effort. Additionally, an adaptive learning law is developed to mitigate the limited performance in the presence of stochastic noise, yielding high convergence accuracy yet without compromising convergence speed. Simulation and experimental comparisons with existing model-based and data-driven inversion-based ILC validate the effectiveness as well as the superiority of the proposed method.