Knowledge-enabled multi-agent deep reinforcement learning approach for batch machine scheduling in semiconductor fabrication

Batch machines are often the scheduling bottleneck in semiconductor fabrication. The presence of complex characteristics, such as batch processing, multi-product mixed processing and re-entrant flow, leads to an unknown mechanism of influence between production factors and performance metrics, which, in turn, prevents global optimization. Therefore, a knowledge-enabled multi-agent deep reinforcement learning (KE-MADRL) approach is proposed. First, a knowledge extraction method based on inverse reinforcement learning is proposed to learn from near-optimal solutions, with which the relationship between key feature parameters and global objective is established. Secondly, a multi-agent deep reinforcement learning model is designed, including the batch forming agent, batch scheduling agent and collaboration mechanisms. The extracted knowledge is incorporated into the agents to guide the design of reward functions. Finally, 24 instances and simulation instances using enterprise data show that the KE-MADRL approach delivers efficient performance, with a 20% reduction in the wafer production cycle time.