Flexciton
Assembly Line
Tapping AI for Leaner, Greener Semiconductor Fab Operations
Flexciton’s software combines two conventional approaches to production scheduling: heuristics and mathematical optimization. “We begin with a relatively simple rule-based algorithm to find a good starting point for our scheduling problem,” Flexciton CEO Jamie Potter told EE Times Europe. “This initial solution is what we call a ‘warm start.’ Then we use mixed-integer linear programming to further optimize the problem from that starting point. However, this problem is often too complex for computers to solve within a practical time frame.
Flexciton uses a cloud-based hybrid scheduling model whose two key components are a global scheduler and a toolset scheduler. The global scheduler focuses on the broader objective of optimizing the overall manufacturing process at the fab level. It considers the big picture and aims to make decisions that maximize efficiency and productivity across the fab. The toolset scheduler works at a more granular level, focusing on the efficient operation of individual machines or tools within the fab. Its main goal is to ensure that wafers move through the fab machines smoothly and efficiently. It also strives to follow the priority ranking that was determined by the global scheduler.
To ensure that even the slightest changes in a fab element’s operation do not hamper the manufacturing process, Flexciton’s software runs its calculations every 10 minutes for the global scheduler and every three minutes for the toolset scheduler.
Could Reinforcement Learning play a part in the future of wafer fab scheduling?
However, as the use of RL for JSS problems is still a novelty, it is not yet at the level of sophistication that the semiconductor industry would require. So far, the approaches can handle standard small problem scenarios but cannot handle flexible problems or batching decisions. Many constraints need to be obeyed in wafer fabs (e.g., timelinks and reticle availability) and it is not easily guaranteed that the agent will adhere to them. The objective set for the agent must be defined ahead of training, which means that any change made afterwards will require a repeat of training before new decisions can be obtained. This is less problematic for solving the instance proposed by Tassel et al., although their approach relies on a specifically modelled reward function which would not easily adapt to changing objectives.