• 목록

[Abstract] 

Yield in semiconductor manufacturing is strongly path-dependent, as lots processed through different machine sequences can achieve substantially different final yields even under comparable throughput conditions. Yet conventional scheduling research has focused almost exclusively on productivity metrics, leaving quality-aware scheduling largely unexplored. This paper proposes a dual Graph Convolutional Network (GCN)-guided Monte Carlo Tree Search (MCTS) framework for yield-aware scheduling in a multistage manufacturing process with parallel machines, where the objective is to maximize product yield while preserving production throughput. The framework integrates two complementary GCN components—a Regression GCN (RegGCN) that learns a path-to-yield value function from historical path-yield records, and a Policy GCN (PolGCN), trained through offline policy distillation from RegGCN, that provides state-conditioned action priors over feasible machine assignments for MCTS under real-time machine availability. These components are integrated into a Progressive Value Integration PUCT (PVI-PUCT) selection mechanism that adaptively blends RegGCN value estimates with simulation statistics as tree search matures, enabling informed decisions from the earliest search iterations. Experiments are conducted on a simulated semiconductor dataset incorporating machine-level main effects and adjacent-stage interaction effects that preclude purely additive yield prediction. Results show that RegGCN achieves an R of 0.9984 using only 5% of all feasible paths, and that the proposed framework improves yield by +11.18 percentage points over a random baseline while maintaining comparable throughput, matching the performance of parameter-tuned heuristics without requiring sensitivity analysis or parameter selection. The analysis further reveals that the GCN value function provides the primary yield-improvement signal, while MCTS lookahead prevents throughput degradation by accounting for future machine availability in early dispatching decisions. These results demonstrate that process-path data can be transformed into actionable scheduling decisions for yield improvement without requiring additional sensors or inline inspection data.