半導體黃光製程（Photolithography）為晶圓製造中的重要瓶頸，其塗佈放行與曝光派工決策會直接影響品質風險與產能表現。晶圓完成塗佈後受到嚴格的等候時間（Queue Time，Q-time）限制，若未能於規定時間內完成後續製程，可能造成重工或報廢；一般檢測不合格所形成的重工回流、首瓶頸站（First Bottleneck，FBN）機制下的母子批等待與合流，以及設備停機，亦會增加在製品負載、等待成本與瓶頸資源競爭，使黃光排程須在Q-time風險、產能與母批等待之間進行動態權衡。本研究運用具動態權重之近端策略優化（Proximal Policy Optimization，PPO），建構可依製程狀態進行塗佈放行、曝光派工與等待決策之單一黃光排程代理人，作為半導體製造業代理式AI決策之探索性實作。研究建立事件驅動模擬環境，納入Q-time限制、一般檢測重工、Q-time超時重工、FBN母子批等待與合流，以及曝光設備停機等機制。Dynamic-PPO依各批次剩餘Q-time寬裕程度與系統負載，逐批調整急迫等待成本，並於每一回合結束後透過自適應懲罰係數反映品質、產能與等待結果。 本研究設置四個單一情境與兩個混合動態日情境，並以20組獨立正式測試種子，比較最小寬裕時間（Minimum Slack Time，MST）、固定急迫倍率PPO（Fixed-PPO）與動態急迫倍率PPO（Dynamic-PPO）之排程表現。結果顯示，Dynamic-PPO於兩個混合動態日均依製程優先序呈現較佳整體結果。其中，在狀態類型、持續時間與設備停機時點皆具隨機性的Mixed-Random-State-Day中，相較Fixed-PPO，Dynamic-PPO將Q-time違規降低69.6%、Q-time重工降低70.1%、產能缺口降低52.4%，並將完成量提高5.6%。研究結果顯示，將批次剩餘寬裕程度與系統負載共同納入Q-time急迫成本，有助於排程代理人因應非平穩生產狀態，並在品質風險控制與瓶頸產能維持之間形成較具調整能力的決策。;Photolithography is a critical bottleneck in semiconductor manufacturing， where coating-release and exposure-dispatching decisions affect both quality and capacity. Queue-time (Q-time) constraints， rework returns， First Bottleneck (FBN) parent–child lot waiting and merging， and equipment breakdowns further increase scheduling complexity. This study applies dynamically weighted Proximal Policy Optimization (PPO) to construct a single photolithography scheduling agent for release， dispatching， and holding decisions as an exploratory implementation of agentic-AI decision making. An event-driven simulation models Q-time constraints， inspection- and Q-time-induced rework， FBN synchronization， and equipment breakdowns. Dynamic-PPO adjusts lot-level urgency costs according to remaining Q-time slack and system workload， while adaptive penalties are updated after each episode based on quality， capacity， and waiting outcomes. Four single-condition and two mixed-day scenarios are evaluated using 20 independent test seeds. Dynamic-PPO is compared with Fixed-PPO and the Minimum Slack Time rule. Under the process-priority evaluation rule， Dynamic-PPO performs better in both mixed-day scenarios. In Mixed-Random-State-Day， it reduces Q-time violations by 69.6%， Q-time rework by 70.1%， and capacity shortfall by 52.4%， while increasing throughput by 5.6% compared with Fixed-PPO. The results indicate that combining lot-level Q-time slack with system workload improves adaptability under non-stationary production conditions while balancing quality risk and bottleneck capacity.
