Scheduling the world's most expensive factories
Semiconductor manufacturing appears on 14 recent papers and wafer fabrication on 3, split between the AI and operations research journals, while chip plants worth tens of billions are being built with public money. Fab scheduling is among the hardest in industry. Do the published methods reach the plants, and how much yield and throughput do they add?
What prompted this
Since October 2025, 48 papers carry a semiconductor or wafer-fabrication keyword (20 in AI journals, 9 in operations research, 8 in economics): "semiconductor manufacturing" on 14, "semiconductor industry" on 12, "wafer fabrication" on 3, "semiconductor supply chain" on 3. Twenty-four of them carry a scheduling, dispatching, production, yield or capacity keyword. The economics of chip subsidies is a separate and larger literature.
Questions
- Do dispatching rules learned by reinforcement learning, as in several recent papers, beat the rules fabs use, on fab-scale simulations with realistic tool failures?
- How much of a new fab's ramp-up time comes from scheduling and yield learning, and can it be shortened by methods from existing plants?
- As fabs multiply across countries with different labour costs, which scheduling constraints change?
Data that could help
Public fab simulation models and datasets (for example the SMT2020 testbed), ramp-up timelines from company reports, and yield data where disclosed.
Next step
Benchmark three recent learned dispatching methods against standard rules on a public fab simulation testbed with tool failures turned on.