A Customizable Simulator for Artificial Intelligence Research to Schedule Semiconductor Fabs

Abstract

Optimal scheduling of semiconductor fabs is a huge challenge due to the problem scale and complexity. New dispatching strategies are usually developed and tested using simulators of different fidelity levels. This work presents a scalable, open-source tool for simulating factories up to real-world size, aiming to support the research into new scheduling algorithms from prototyping to large-scale experiments. The simulator comes with a declarative environment definition framework and is out of the box usable with existing reinforcement learning methods, priority-based rules, or evolutionary algorithms. We verify our tool on large-scale public instances and provide proof-of-concept demonstrations of the reinforcement learning interface’s usage.