An Inherently Explainable Approach for Reinforcement Learning Based Dispatching in Semiconductor Frontend Fabs

Scheduling is a crucial task in many manufacturing environments. Depending on the type of the manufactured product, a multitude of different properties and constraints complicate the already complex structure of a job-shop scheduling problem. Semiconductor Manufacturing is well known to be one of the most complex production processes, namely for its stochastic nature, diverse product mix, which leads to strict tool dedications, and for the presence of re-entrant flows in the production line. In this paper, we propose an approach for finding a dispatching policy through the use of (NN), whose weights are optimized through an Evolution Strategies method called (CMA-ES), able to minimize the tardiness, throughput and other relevant metrics within a digital twin of a real-world, stochastic, large-scale semiconductor manufacturing facility.