摘要
This study investigates the semiconductor backend assembly scheduling problem, a critical challenge in make-to-order manufacturing where minimizing flow time is essential. The problem features identical and unrelated parallel machines, setup times, and entity transformations (e.g., wafers diced into dies, dies bonded to substrates, substrates packed into magazines). Additional complexities—including batch formation, job splitting, and machine eligibility—further increased the difficulty, rendering it an NP-hard extension of the hybrid flow shop problem. Traditional heuristic algorithms and simulation based optimization often face scalability and efficiency limitations in large-scale applications. To address these challenges, we propose an integrated framework that combines deep reinforcement learning with simulation-based optimization, supported by simplified swarm optimization and an optimal replication allocation strategy. Simplified swarm optimization explores the solution space, while optimal replication allocation directs simulation resources towards promising solutions, enhancing efficiency. Deep reinforcement learning strengthens the framework by dynamically adjusting dispatching rules, accelerating decision-making, and avoiding local optima. Experiments on 18 datasets from a Taiwanese semiconductor factory demonstrated that the proposed approach substantially reduces average flow time while preserving computational efficiency, offering an effective solution to this complex scheduling problem. © 2025 Elsevier B.V.