Deep reinforcement learning for workload balance and due date control in wafer fabs

Semiconductor wafer fabrication facilities (wafer fabs) often prioritize two operational objectives: work-in-process (WIP) and due date. WIP-oriented and due date-oriented dispatching rules are two commonly used methods to achieve workload balance and on-time delivery, respectively. However, it often requires sophisticated heuristics to achieve both objectives simultaneously. In this paper, we propose a novel approach using deep-Q-network reinforcement learning (DRL) for dispatching in wafer fabs. The DRL approach differs from traditional dispatching methods by using dispatch agents at work-centers to observe state changes in the wafer fabs. The agents train their deep-Q-networks by taking the states as inputs, allowing them to select the most appropriate dispatch action. Additionally, the reward function is integrated with workload and due date information on both local and global levels. Compared to the traditional WIP and due date-oriented rules, as well as heuristics-based rule in literature, the DRL approach is able to produce better global performance with regard to workload balance and on-time delivery.