| 摘要: | 研究中探索了機器學習在提高非富勒烯三元有機太陽能電池（Ternary Organic Solar cell, TOSC）開路電壓（open-circuit voltage , VOC）方面的潛力。為了解決TOSC 中低 VOC 阻礙功率轉換效率進一步提高的挑戰。收集了 2016 年至 2023 年的 407 個實驗數據點，並專注於非富勒烯體異質接面 TOSC的主動層。 VOC 值的高斯分佈顯示集中在 0.875 V 左右，突顯了能階匹配以實現 VOC 以上 0.95 V 的必要性。
採用極限梯度提升 (XGBoost) 和人工神經網路 (ANN) 兩種機器學習模型，使用供體、受體 1 和受體 2的能階及三者組成的質量百分濃度比作為描述子來預測 VOC。 ANN 模型表現出優越的性能，驗證和測試相關係數分別為 0.84 和 0.88，並且其強大的外推能力也得到了驗證。 SHAP值分析確定A1_LUMO是影響VOC預測的關鍵因素。
為了評估外推法，我們定義了一個包含前 5% VOC 值的外推測試集。 ANN/EXP 模型的精確度為 75%，召回率為 43%，F1 分數為 55%，顯著優於 XGBoost/EXP 模型，後者未能推斷出非凡的 VOC 系統。 ANN/DFT 模型展示了顯著的外推能力，實現了 100% 的精度、43% 的召回率和 60% 的 F1 分數。這些結果強調了 ANN 模型在預測異常 VOC 值方面的卓越表現。
對222,374個TOSC材料重組進行高通量虛擬篩選，鑑定出9,742個預測VOC > 0.95 V的組合和1,398個> 1 V的組合，並使用密度泛函理論 (DFT) 描述子對哈佛乾淨能源計畫資料庫 (HCEPDB) 預測後的進一步片段分析證實了能階匹配在優化 VOC 方面的重要性。
這項研究強調了具有強大外推能力的 ANN 模型的潛力，可以透過預測最佳材料組合和能階範圍（甚至對於不可見的數據）來指導高效 TOSC 的設計。這種方法有望加速先進有機光伏材料的開發，在克服當前 VOC 限制和提高太陽能電池整體效率方面邁出重要一步。
;This study investigates the potential of machine learning (ML) to enhance the open-circuit voltage (VOC) in non-fullerene ternary organic solar cells (TOSCs). We addressed the challenge of low VOC, which impedes further improvements in power conversion efficiency (PCE), by collecting 407 experimental data points from 2016 to 2023, focusing on the active layers of non-fullerene TOSCs. The VOC values′ Gaussian distribution centers around 0.875 V, highlighting the need for matched energy levels to achieve VOC above 0.95 V.
Two ML models, Extreme Gradient Boosting (XGBoost) and Artificial Neural Network (ANN), were employed to predict VOC using energy level descriptors and mass percentage concentration ratio of donor, acceptor 1, and acceptor 2 components. The ANN model demonstrated superior performance, with validation and test correlation coefficients of 0.84 and 0.88, respectively, and was verified for its robust extrapolative capability. SHAP value analysis identified A1_LUMO as the key factor influencing VOC prediction.
To assess extrapolation, we defined an extrapolated test set comprising the top 5% of VOC values. The ANN/EXP model achieved a precision of 75%, a recall of 43%, and an F1 score of 55%, significantly outperforming the XGBoost/EXP model, which failed to extrapolate extraordinary VOC systems. The ANN/DFT model demonstrated significant extrapolation capabilities, achieving a precision of 100%, a recall of 43%, and an F1 score of 60%. These results underscore the superior performance of ANN models in predicting extraordinary VOC values.
High-throughput virtual screening of 222,374 TOSC material combinations identified 9,742 combinations with predicted VOC > 0.95 V and 1,398 combinations > 1 V. Optimal energy level ranges for high VOC prediction were established, and Density Functional Theory (DFT) calculations on the Harvard Clean Energy Project Database (HCEPDB) further validated the importance of energy level matching.
This research highlights the potential of ANN models with strong extrapolative capabilities to guide the design of high-efficiency TOSCs by predicting optimal material combinations and energy level ranges, even for unseen data. This approach promises to accelerate the development of advanced organic photovoltaic materials, offering a significant step forward in overcoming current VOC limitations and enhancing overall solar cell efficiency. |
