AI Solves Complex Factory Scheduling 25% Better

In semiconductor manufacturing, where production efficiency directly impacts global technology supply, a persistent has been scheduling hundreds of jobs across multiple machines while minimizing costly setup times. This problem, known as serial batch scheduling with minimum batch sizes, appears in critical areas like photolithography and ion implantation, where changing machine configurations between different product families requires expensive downtime. Researchers have now developed a new constraint programming approach that significantly outperforms existing s, finding solutions up to 25% better for large-scale instances with up to 500 jobs, 10 families, and 10 machines.

The key breakthrough comes from a novel Aligned constraint programming model that eliminates the need for a predefined virtual set of possible batches, which had previously limited scalability. Instead of creating artificial batch indices that exponentially increase problem complexity, the new reasons directly on sequences of same-family jobs scheduled on machines. This allows the model to construct what the researchers call "family blocks"—consecutive jobs from the same family that can be processed without setup changes—through a more compact formulation that avoids the curse of dimensionality plaguing earlier approaches.

Ology builds on three interconnected sections that work together to find optimal schedules. First, a Core section sequences jobs on machines while respecting release times and setup requirements between different families. Second, a Family Block section introduces virtual interval variables that extend over the entire duration of each family block, aligning jobs from the same family. Third, a Sizing section uses cumulative functions over these aligned virtual intervals to enforce minimum and maximum batch size requirements. The researchers further enhanced this approach with tailored search phases that first assign core scheduling decisions before handling the auxiliary virtual intervals, creating an improved model they call s-A*.

Extensive computational experiments on nearly five thousand instances demonstrate the superiority of the new approach. On small-to-medium instances with up to 100 jobs, the proposed models consistently outperformed existing mixed-integer programming formulations and earlier constraint programming models, achieving improvements of more than 7% over the RP mixed-integer model within just 10 minutes. More significantly, on large-scale instances with up to 500 jobs, the new models found solutions up to 25% better than the best existing constraint programming approaches, with average percentage improvements steadily increasing as instance size grew from 100 to 500 jobs. The researchers tested their models against seven existing approaches across 3,750 large-scale instances, with the s-A* model showing particular strength on the most challenging problems.

For semiconductor manufacturing are substantial, as can generate high-quality schedules in under 20 minutes for realistic wafer-fab scale problems. This addresses a critical need in environments where scheduling tools typically regenerate plans every 5 to 10 minutes. The approach specifically helps with ion implantation operations, where changing implant gases is costly and time-consuming, and photolithography areas, where rare gas availability constraints make efficient batching essential. By ensuring minimum batch sizes are met while minimizing total weighted completion time, balances operational constraints with production efficiency in ways previous s could not achieve at scale.

Despite these advances, the researchers acknowledge limitations in their current implementation. The most challenging instances for the proposed models are those where the number of families equals the number of machines, as increased symmetries in these cases make finding optimal solutions more difficult. The models also do not solve every large-scale instance within the 20-minute time limit, finding solutions on approximately 76% of the most challenging problems. Additionally, while shows significant improvements over existing approaches, incorporating symmetry-breaking constraints without pruning potentially optimal solutions remains an area for future research to further enhance performance on symmetric instances.