Deep Learning Agents for Efficient Dynamic Production Control in Semiconductor Manufacturing

This research utilizes deep learning (DL) agents for efficient dynamic production control in large complex manufacturing systems, such as semiconductor manufacturing. The DL agents can rapidly generate production control policies and perform immediate re-calculation whenever changes occur in the shop floor, such as when product-mix changes, machines are replaced, or machines are qualified for production. Since in the advanced process control of semiconductor manufacturing, it is common to have machines focusing on a small number of product types, we design the DL agents to solve this problem size and then replicate their use for all machines in any workstation. To explicitly consider random occurrences in the shop floor, the DL agents are trained on a large dataset of production control policies that are optimally solved using a Markov decision process (MDP) formulation. Numerical experiments show that the DL agents can generate near-optimal production control policies in less than 1% of the computational time of MDPs while achieving 95% accuracy in terms of correct decisions. Simulations show that the framework improves cycle time by over 90% compared with traditional dispatching rules.