AI-driven optimization and digital twin simulation for robust, scalable production scheduling

Production environments with high variability and frequent disruption need plans that are good, that hold up under disturbance, and that stay current as information arrives. Optimisation, simulation and AI each address these needs in depth, but separately. This thesis integrates them into a single, innovative AI-driven scheduling framework, developed and evaluated on the mould validation and testing laboratory of Novo Nordisk. An optimisation layer combines an exact solver, a dispatching heuristic and a genetic algorithm to generate a pool of distinct candidate plans. A discrete-event simulation ranks them under stochastic disturbance, to reproduce the unpredictability of the real world. A large language model converts operational information into parameter changes automatically, as soon as it arrives, and triggers a reschedule. A planner approves every new plan. The results show that nominal quality predicts execution only weakly. The role of the solvers is therefore not to deliver the most optimised plan nominally, but many good and different candidates, among which the simulation finds the one that performs best under real-world disturbances. When resources are short, the simulation also reveals lateness that the nominal plan hides. The language model reads operational information reliably and traceably, with no wrong decision in 2,000 readings. In the end-to-end test, an operational event became a measurably better plan in about half an hour. Optimisation, simulation and AI-driven rescheduling can therefore be integrated into one framework that meets the needs of the laboratory: the plan to issue is chosen by its simulated behaviour, and kept current as events arrive. The framework is now in pilot at the company.