高效率有機太陽能電池（OPVs）的開發面臨重大挑戰，主要源於供體：受 體：第三元件（D:A:T）材料組合所構成的龐大組合空間。本研究提出一套具備通用性與可擴展性的片段式圖神經深度學習框架，用以加速非富勒烯（NF）三元OPV材料的預測設計與發現。我們將分子表示為由具化學意義的功能性片段（FF）組成的圖結構，並結合來自密度泛函理論（DFT）與分子力學所計算之電子與結構性片段描述子，訓練一個導向訊息傳遞神經網路（D-MPNN）模型以精準預測光電轉換效率（PCE）。為了提升高效能系統的預測準確度，我們引入加權訓練策略，對 PCE > 16% 的樣本賦予更高權重，使模型在高效率區間的預測表現顯著改善。加權後的 D-MPNN 模型於交叉驗證中達到優異表現（r = 0.93、MAE = 0.89%、RMSE = 1.12%），並能有效泛化至時間上獨立的測試資料集。我們進一步應用該模型篩選超過 3.34 億種 D:A:T 組合，成功辨識出 153 組預測 PCE 超過 20% 的高效率系統。透過 Z-score 增富分析，我們辨識出對高效能表現有顯著貢獻的關鍵分子片段，為未來 OPV 材料設計提供資料驅動的指引。本片段式框架提供了一條穩健、高效且具可擴展性的高通量虛擬篩選管線，能有效支援有機太陽能材料的加速發現。;The development of high-efficiency organic photovoltaics (OPVs) remains a significant challenge due to the vast combinatorial space of donor: acceptor: ternary (D: A: T) material systems. In this work, we present a generalizable and scalable fragmentbased graph deep learning framework to accelerate the predictive design and discovery of non-fullerene (NF) ternary OPV materials. By representing molecules as graphs composed of chemically meaningful functional fragments (FFs), and encoding fragment-level electronic and structural descriptors derived from density functional theory (DFT) and molecular mechanics, we train a directed message passing neural network (D-MPNN) to accurately predict power conversion efficiency (PCE). To enhance prioritization of high-performance systems, we implemented a weighted training scheme that assigns greater importance to samples with PCE > 16%, improving predictive accuracy in the high-efficiency regime. The weighted D-MPNN achieves excellent performance (r = 0.93, MAE = 0.89%, RMSE = 1.12%) in cross-validation and generalizes effectively on a temporally separated independent test set. We apply the model to screen over 334 million D:A:T combinations, identifying 153 systems with predicted PCEs exceeding 20%. Enrichment analysis using Z-scores reveals key molecular fragments contributing to high performance, offering data-driven insights for future OPV design. This fragment-based approach provides a robust, efficient, and extensible pipeline for high-throughput virtual screening in OPV research.
