| 摘要: | 在Micro-LED巨量轉移技術以雷射誘導前向轉移技術(LIFT)最具代表性，因其有非接觸、高速與高精度等優勢，被視為實現高效製程的潛力方案。然而，當涉及晶片品質分選與放置間距調整等需求時，機台頻繁地移動將顯著降低效率並增加成本負擔。本研究針對Micro-LED巨量轉移製程中因晶片品質分布差異、間距調整及晶圓更換所衍生的挑戰，提出一套巨量轉移智慧系統。基於強化學習架構，藉由自主決策與經驗累積，同時滿足品質揀選與間距調整需求。相較於傳統最佳化方法，強化學習能透過互動式學習獲取最適策略，並於獎勵函數整合多目標展現動態適應性，突破LIFT技術在複雜轉移條件下的效能限制。研究中將巨量轉移任務建模為馬可夫決策過程(MDP)，定義狀態(State)、動作(Action)與獎勵(Reward)，並以最小化轉移時間為優化目標。巨量轉移智慧系統包含LIFT技術模擬環境、快速匹配掃描演算法與多種強化學習整合與比較，並接續導入函數近似之深度學習與遷移學習機制，以提升跨晶圓樣態的泛化能力。最終，代理能自主學習高效率的轉移路徑規劃策略，有效縮減整體製程時間，展現本研究架構於產業實務中的可行性與應用潛力。
為優化狀態建構效率，本研究設計向量化批次搜尋法，模擬結果顯示，晶片數達500個以上時，較傳統拉鍊式逐位掃描法可節省高達94.51%的CPU時間。巨量轉移智慧系統以126顆來自真實晶圓之數據進行驗證，在同時考量品質分選與間距調整之條件下，逐顆轉移方法耗時165.2秒，產能僅2745 UPH；而基於Q-Learning架構之結果僅需54.53秒，產能達8314 UPH，轉移耗時縮減67%，整體效率提升至3.03倍，驗證其於複雜轉移需求下之可行性與效能提升。後續，本研究導入深度強化學習架構，以克服傳統Q表在高維度狀態與動作空間下的儲存限制與泛化不足問題。採用DQN作為模型主體，透過神經網路的非線性特徵擷取能力學習狀態與動作間的複雜關係，提升策略學習於高維空間下的表現。同時提出基於樣本價值評估的經驗池更新策略，動態調整樣本保留機率以強化高價值樣本，並剔除冗餘或獎勵衝突的樣本，有效降低劣質經驗對訓練的干擾。以126顆真實晶圓數據進行模擬，DQN架構僅需51.9秒即可完成任務，產能達8727 UPH，轉移耗時縮減68.6%，整體效率提升3.18倍，展現本研究於複雜轉移需求下的顯著效能與實務潛力。
為驗證在不同晶圓資料下的延續性與再訓練能力，使用遷移學習將原先訓練完成之模型參數與架構應用於新晶圓透過微調法(Fine-tuning)接續訓練，無需重新初始化參數。結果顯示，遷移學習相較於獨立訓練在多項指標上均具顯著優勢。使用126點變異指標D_wafer為0.05之數據驗證，獨立訓練需約550次更新才能完成一回合，以部分遷移為例，透過遷移學習僅約412次，初始即可減少138次更新，訓練效率提升約25.5%；在收斂表現上，部分遷移學習約於第1200回合即降至138次以下並保持穩定，較獨立訓練提前約800回合，收斂速度提升約40%，充分展現跨晶圓任務中的效率優勢與成本節省潛力。此外，本研究提出的晶圓變異指標D_wafer驗證遷移學習的實務適用性。當D_wafer≤0.1時，模型可直接沿用基準晶圓成果，快速收斂並維持效能，降低再訓練成本並提升跨晶圓適應性與部署效率。
本研究所提出之巨量轉移智慧系統為少數針對LIFT巨量轉移過程中同時考量品質分選與間距調整需求所設計之演算法。該方法除可提升平台移動效率與轉移速度，亦於模型結構設計、樣本選擇機制與再訓練適應性方面進行整體優化，具備高效能與產線應用潛力，對推動Micro-LED巨量轉移之實務化與智慧製造具有實質貢獻。
;In Micro-LED mass transfer technology, laser-induced forward transfer, LIFT, is regarded as the most representative approach due to its non-contact, high-speed, and high-precision characteristics, and is widely considered a potential solution for achieving high-efficiency manufacturing. However, when chip quality binning and placement pitch adjustment are required, frequent equipment movement significantly reduces throughput and increases production cost. To address the challenges arising from chip quality variation, spacing adjustment, and wafer replacement, this study proposes an intelligent mass transfer system. Built upon a reinforcement learning framework, the system enables autonomous decision-making and knowledge accumulation to simultaneously satisfy quality binning and pitch adjustment requirements. By formulating the mass transfer process as a Markov Decision Process with well-defined states, actions, and rewards, the objective is set to minimize overall transfer time. The proposed system integrates an LIFT-based simulation environment, a fast matching algorithm, and multiple reinforcement learning models, and is further extended with deep reinforcement learning and transfer learning mechanisms to enhance generalization across different wafer distributions.
To improve state-construction efficiency, a Vectorized Batch Search Method, VBSM, was designed, which demonstrated up to 94.51 percent CPU time savings compared with the traditional Zigzag Scanning Method when processing over 500 chips. Experimental validation using 126 real wafer data points shows that the conventional die-by-die method required 165.2 seconds with a throughput of only 2745 UPH, whereas the Q-Learning–based system completed the same task in 54.53 seconds, achieving 8314 UPH, a 67 percent reduction in transfer time and a 3.03 times efficiency improvement. Subsequently, a deep Q-network, DQN, framework was introduced to overcome the storage and generalization limitations of traditional Q-tables under high-dimensional state-action spaces. With the nonlinear feature extraction of neural networks, the DQN system effectively captured complex state–action relationships. An experience replay strategy based on sample value evaluation was also proposed, dynamically adjusting retention probabilities to emphasize high-value samples and discard redundant or conflicting experiences, thereby reducing instability during multi-objective training. Using 126 real wafer data points, the DQN-based system completed the same task in 51.9 seconds with a throughput of 8727 UPH, reducing transfer time by 68.6 percent and improving efficiency by 3.18 times, demonstrating its practical feasibility and robustness under complex transfer requirements.
Furthermore, transfer learning was adopted to evaluate the continuity and adaptability of trained models across different wafers. Rather than retraining from scratch, the pretrained model was adapted to new wafer data through fine-tuning. The results indicate that transfer learning consistently outperforms independent training. With 126-point wafer variation data where D_wafer equals 0.05, independent training required approximately 550 updates per episode, whereas transfer learning reduced this to 412 updates, lowering training load by 138 updates, a 25.5 percent improvement. In terms of convergence, partial transfer learning achieved stable performance at around 1200 episodes, approximately 800 episodes earlier than independent training, representing a 40 percent faster convergence; sequential transfer achieved stability by 1500 episodes, about 25 percent faster. Moreover, the proposed wafer variation index D_wafer was validated as a reliable measure for practical applicability. When D_wafer is less than or equal to 0.1, pretrained models could be directly reused, enabling rapid convergence with performance comparable to baseline wafers, thereby reducing retraining cost and enhancing deployment efficiency.
Overall, the proposed intelligent mass transfer system is among the few approaches that simultaneously address quality binning and pitch adjustment requirements within LIFT-based mass transfer processes. By optimizing platform motion efficiency, transfer speed, model design, sample selection, and retraining adaptability, this study demonstrates strong performance, scalability, and industrial applicability, offering substantial contributions to the practical implementation of Micro-LED mass transfer and advancing smart manufacturing. |
