Scholarlink

<p>Due to the uncertainties and complex disturbances associated with semiconductor manufacturing processes, fixed-gain sliding mode observer-based run-to-run (SMO-RtR) control struggles to achieve superior performance. To this end, we propose an enhanced control framework by integrating deep reinforcement learning (DRL) into SMO-RtR, where the well-trained DRL agent generates the optimal policy to provide suitable gain for the SMO, and the optimized SMO-RtR yields a superior control recipe for the semiconductor manufacturing process. First, a detailed stability and convergence analysis of the SMO-RtR controller is conducted under complex compound disturbances. Then, the gain optimization problem of SMO is modeled as a Markov decision process, where a Huber function-based reward is designed to guide the agent to learn the manufacturing environment. In addition, a new DRL agent that integrates structured control nets (SCN) and the stochastic weighted twin-delayed deep deterministic policy gradient (SWTD3) algorithm is presented to improve the policy optimization and decision-making capabilities. To evaluate the effectiveness of the suggested approach, an implementation called SCN-SWTD3-SMO is developed. Experimental results demonstrate that DRL agents can adaptively adjust the gain of the SMO in response to changing disturbances, thereby improving the anti-disturbance and tracking performance. The findings underscore the potential application of DRL techniques to enhance the performance of RtR control in semiconductor manufacturing processes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intelligent run-to-run control of semiconductor manufacturing processes using deep reinforcement learning enabled sliding mode observer

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Zhu Ma,
  • Yujian Yang,
  • Qiang Wang,
  • Tianhong Pan

摘要

Due to the uncertainties and complex disturbances associated with semiconductor manufacturing processes, fixed-gain sliding mode observer-based run-to-run (SMO-RtR) control struggles to achieve superior performance. To this end, we propose an enhanced control framework by integrating deep reinforcement learning (DRL) into SMO-RtR, where the well-trained DRL agent generates the optimal policy to provide suitable gain for the SMO, and the optimized SMO-RtR yields a superior control recipe for the semiconductor manufacturing process. First, a detailed stability and convergence analysis of the SMO-RtR controller is conducted under complex compound disturbances. Then, the gain optimization problem of SMO is modeled as a Markov decision process, where a Huber function-based reward is designed to guide the agent to learn the manufacturing environment. In addition, a new DRL agent that integrates structured control nets (SCN) and the stochastic weighted twin-delayed deep deterministic policy gradient (SWTD3) algorithm is presented to improve the policy optimization and decision-making capabilities. To evaluate the effectiveness of the suggested approach, an implementation called SCN-SWTD3-SMO is developed. Experimental results demonstrate that DRL agents can adaptively adjust the gain of the SMO in response to changing disturbances, thereby improving the anti-disturbance and tracking performance. The findings underscore the potential application of DRL techniques to enhance the performance of RtR control in semiconductor manufacturing processes.