Cyclic reinforcement learning for generalization enhancement on T/C imbalance scheduling in TFT-LCD cell manufacturing

Abstract

The rising product diversity for Thin-Film Transistor Liquid Crystal Display (TFT-LCD) has amplified the need for an efficient manufacturing process. This study formulates the TFT-LCD cell process scheduling as a dynamic flexible job shop scheduling problem, aiming to balance production between TFT array and color filter substrates (i.e. T/C balance) while accounting for new job arrivals and uncertain processing times. To optimize multiple objectives, including makespan, total weighted tardiness, violation of limited queue time, and T/C balance, a cyclic reinforcement learning (CRL) framework with a cyclic training process is proposed to achieve robustness under uncertain scenarios. A numerical study is conducted to validate the proposed framework, with performance compared against benchmark models, including optimization-based approaches and genetic algorithm. The results show that the CRL outperforms benchmark models in both realized objective value and variation while efficiently handling new job arrivals within a short inference time. Sensitivity analysis further confirms the robustness even in highly uncertain manufacturing environments.