(698a) On State Estimation in the Low-Volume High-MIX Semiconductor Manufacturing Using a Singular Gauss-Markov Model

The need for advanced process control (APC) in the semiconductor manufacturing industry has been well recognized and the benefits have been demonstrated in many applications such as polishing, etch and litho processes. Although feed-forward control is implemented in several processes recently, run-to-run feedback control has been the most commonly used control technique for different processes. The basic structure of a run-to-run controller consists of a process model, a state estimator and a controller, where different process models can be applied for different applications to transform the output measurements into states, the state estimator estimates the state based on the previous estimate and the current measurement, then the controller calculates the control move based on the estimated process state for the next run. Because linear process model and dead beat control law are applied in most applications, state estimation method becomes the key component that has significant impact on the control performance.

In order to generate an accurate estimated state, the state estimator should be able to track the process changes and to reject the measurement noises simultaneously. There are many sources of the process variations, such as tools, devices, layers etc. Because there is no direct measurement of each individual source, the traditional method is to segregate the historical data into different bins and use the ones that match the current event of the specific tool, device and layer, etc. for feedback control. This segregation is based on the idea that wafers with similar process history have similar characteristics, and the control based on the segregated data is also known as ?threaded? control. The benefit of threaded control is that within the thread definition, one does not need to worry about individual state changes because the overall state will track all of them as a lumped parameter; while the disadvantage is that this narrowly defined process stream can result in insufficient data and a large amount of ?send-ahead? wafers [14] when applied to low-volume high-mix manufacturing, especially for litho processes where many factors contribute significantly to the process variations.

Because many state-of-the-art fabs are operating with increasingly diversified product mixes, run-to-run state estimation methods in low-volume high-mix production have drawn considerable interest in the last few years, and several different methods have been reported [1-9]. Some of the methods have certain theoretical background, such as [1, 4], while others are largely ad hoc, such as [3, 7]. Among these methods, the general theme is trying to share information among different contexts, and different algorithms such as regression and Kalman filter are applied to sort out the contribution from different variation sources.

In this work, a theorem is first formulated and rigorous proof is given to show that for high-mix production the matrix that contains the context information is always rank deficient, therefore, unbiased estimates exist only for certain linear combinations of the unknown parameters (i.e. the linear combinations of the individual state of each component). Then a general singular Gauss-Markov model is used to represent the high-mix production processes. Based on this model, the best linear unbiased estimate (BLUE) [10-11] of the system states is derived and extended to a unifying framework, where the effect of initial estimates of individual state and associated error variance are explicitly represented. In this framework, the published analytical algorithms [1, 3] can be explained as special cases. Using this framework, the advantage and disadvantage of different algorithms are discussed and a new state estimation method is proposed. In the proposed method, recursive least squares with adaptive forgetting factor is applied to track system states, and a numerically stable algorithm [12] is applied to handle the singular Gauss-Markov model. Process changes are detected by a Bayesian algorithm [13], and the forgetting factor is adjusted once the process change is detected. Simulation and industrial examples are given to illustrate the performance of the proposed algorithm.

Key words: low-volume high-mix, regression, Gauss-Markov process, best linear unbiased estimate (BLUE), Kalman filter, threaded control.

Reference 1. A.J. Toprac and Y. Wang (2005), Advanced method for run-to-run control of photolithography overlay registration in high-mix semiconductor production, in Proceedings of SPIE - data analysis and modeling for process control II, vol. 5755, pp1-8. 2. T. Levin, I. Geier, A. Zhivotovsky, N. Aframiam, H. Friedlander-Klar (2005), Automated process control optimization to control low volume products based on high volume products data, in Proceedings of SPIE - data analysis and modeling for process control II, vol. 5755, pp145-156. 3. E. Mos, V. Wang, A. Kisteman, L. Verstappen and H. Megens (2005), Generalized data-sharing and frequency domain optimization of overlay run-to-run feedback control, in 6th European AEC/APC Conference, Dublin, Ireland, March 2005. 4. C.K. Hanish (2005), Run-to-run state estimation in systems with unobservable states, in AEC/APC symposium XVII, Indian Wells, CA, Sep. 24-29, 2005 5. SS. Jiang and D. SH Wong (2005), Stability and Performance Analysis of Mixed Product Run-To-Run Control, in AEC/APC Symposium-Asia, Hsin Chu, Taiwan, Dec. 2005. 6. H. Yue (2005), Run-to-run exposure controller and fault detection in a high-mix development and manufacturing, in AEC/APC symposium XVII, Indian Wells, CA, Sep. 24-29, 2005. 7. EnvironmentT. Chowdhury, M. Freeland, O. Krogh, G. Narasimhan and G. Raghavendra (2004), Propagation of APC models across product boundaries, in Proceedings of SPIE - data analysis and modeling for process control, vol. 5378, pp 74-80. 8. X. Wan, A. Zhou, F. Zhang, J. Li, X. Gu, E. Mos, A. Kisteman, V. Wang, and R. Schuurhuis (2004), Overlay advanced process control for foundry application, in Proceedings of SPIE - Metrology, Inspection, and Process Control for Microlithography XVIII, v 5375, n PART 2, pp 735-743. 9. T. Conway, A. Carlson and D. Crow (2003), Improving overlay control through proper use of multi-level query APC, in Proceedings of SPIE - Advanced Process Control and Automation, vol. 5044, pp24-31. 10. H. J. Werner (1985), More on BLU estimation in regression models with possibly singular covariance, Linear algebra and its application, vol. 67, pp 207-214. 11. A. Albert (1973), The Gauss-Markov theorem for regression models with possibly singular covariance, SIAM journal of applied mathmatics, vol. 24(2), pp 182-187. 12. M. Morf and T. Kailath (1975), Square-root algorithms for least squares estimation. IEEE Transactions on Automatic Control, AC-20: 487-497. 13. J. Wang and Q.P. He (2005), A pattern matching approach for fast disturbance detection and classification using Bayesian statistics, in AEC/APC symposium XVII, Indian Wells, CA, Sep. 24-29, 2005. 14. ?send-ahead? wafers are test wafers to initialize the estimate of a new thread, as known as ?child lots? or ?pilot? wafers.