Generative AI for Production Scheduling: A Graph Diffusion Model for the Flexible Job Shop

The Flexible Job Shop Scheduling Problem (FJSP) is central to modern manufacturing, where high-quality schedules must often be found under a limited budget of expensive, simulation-based evaluations. In this setting, metaheuristics such as genetic algorithms reach strong solutions but consume many evaluations, whereas standard reinforcement-learning-based methods build the schedule decision by decision, reaching a solution quickly but with no way to revise or improve it once produced. This paper proposes SILDiff (Simulator-In-the-Loop Diffusion), a graph-based diffusion framework that uses a learned generative prior to quickly produce a complete schedule of machine assignments and operation sequencing, and can then revise and refine its own decisions. At selected steps during generation, the emerging schedule is evaluated by a discrete-event simulator, improved by a fast local search, and fed back into the denoising process. This gives the diffusion model the corrective power of improvement-based search while retaining the low online cost of its learned generation. Experiments show that, on in-distribution instances, SILDiff matches or exceeds the solution quality of genetic-algorithm and reinforcement-learning baselines while requiring substantially fewer online evaluations. This quality advantage is retained on out-of-distribution instances at roughly twice the training scale.