Performance Assessment For The Run-to-Run Controller

The semiconductor manufacturing industry has been an important strategic position in the national economy and social development as well as national security, which promotes the development of information industry: communications, computers, information appliances and network technology. So, the performance of controller is directly related to the quality of wafer fabrication. However, the performance of controller will degrade when time goes on. It is important to add performance monitoring and evaluation unit into the control system, which ensure process safe, efficiency, low-power consumption. In this paper, the performance assessment algorithm of Run-to-Run(R2R) control based on least squares and Bayesian theory is proposed to solve the problems of interference suppression. The main works are follows:1. The plant-model mismatch phenomenon always exists in semiconductor manufacturing processes. To improve the performance of R2 R controller, a reachable criterion is proposed to assess the performance of R2 R controller using the least squares identification algorithm. Firstly, the variance that impacts on performance of R2 R controller is analyzed, and a formulation of reachable criterion is proposed. Then the parameters are estimated by using the least squares algorithm and rolling time window. Finally, simulations on thread-EWMA controller and double-EWMA controller are validated the proposed algorithm.2. To reduce the influence of diverse disturbances, a performance assessment criterion of Bayesian posterior probability is proposed in this paper, which combines the least squares identification, ARMAX model and Bayesian theory. Firstly, the formulation of disturbances under different situation is presented and the ARMAX model is also built. If the R2 R controller’s performance deteriorates, there is a drift. To capture the drift, the input/output data is divided into two adjacent moving windows according to the manufacturing sequence. Then, the drift occurrence probability is determined based on the value of Bayesian posterior probability in the second moving window. The probability value is used to judge whether the controller performance has a degradation. The influence of disturbances(i.e. plant model mismatch, incorrect tuning, metrology delay, high-mix product) on performance assessment criterion is also analyzed. Finally, the simulation demonstrates the effectiveness of the proposed criterion.