A multi-population cooperative Jaya algorithm with adaptive rescheduling strategy for dynamic scheduling in semiconductor manufacturing system

Time-window constraints (TWCs) and dynamic disruptions complicate semiconductor wafer fabrication scheduling. To address this issue, we present a rescheduling approach incorporating deep reinforcement learning and the multi-population Jaya algorithm (DRL-MPJA). An adaptive DRL-based rescheduling trigger handles uncertain events, and the DRL agent adaptively selects the optimal strategy by real-time states and rewards. For complete rescheduling, three sub-populations optimise makespan, schedule deviation, and TWC feasibility, forming a reference population and guiding sequence. GEPA provides unified evaluation, while GEPRG selects promising candidates based on similarity, enabling cooperative optimisation. Experiments on 12 instances and four-month wafer-fab simulation demonstrate the approach’s superiority.