應用深度學習於Cu2O/Si光偵測器陣列光電流訊號之快速波長辨識

Abstract

傳統光譜辨識系統仰賴外部濾光片、光柵或光譜儀，雖具良好解析能力，但受限於體積、成本與整合複雜度，不利於可攜式與即時感測應用。因此，本研究以溶液製程製備之 4 × 4 p-Cu2O/n-Si 自供電光偵測器陣列為平台，結合深度學習，利用不同波長所造成之暫態光電響應差異，建立無濾光片智慧波長辨識架構。 本研究首先導入 LSTM 與 BiLSTM 模型，以截短光電流訊號辨識四種波長。結果顯示，BiLSTM 在 64 個隱藏單元下，僅使用原始 150 ms 訊號之前 40 ms，即可達到 100% 分類準確率，證明早期暫態訊號已包含足以辨識波長之關鍵資訊。 其次，為提升模型效率與可解釋性，本研究導入 self-attention-enhanced MLP，使模型自動聚焦於具高度判別力之時間區段。該模型於完整訊號下達 95.73% 準確率，且僅使用縮短至 20 ms 資料集時仍維持 94.44% 準確率。結果顯示，模型可有效擷取暫態光電響應中最具判別力的時間區段，進而降低對完整量測訊號的依賴。 最後，針對單一光電流訊號易產生光譜混疊之問題，本研究提出 Isc/Voc 雙通道辨識策略，分別擷取載子傳輸與電荷累積相關特徵。將雙通道暫態訊號整合至 LSTM 後，模型於紫外至近紅外波段達到 98.72% 分類準確率，有效降低相近波長間之辨識混淆。 綜合而言，本研究證實，p-Cu2O/n-Si 自供電光偵測器陣列結合截短訊號分析、注意力機制與 Isc/Voc 雙通道融合，可實現快速、高準確率且具物理可解釋性之無濾光片波長辨識。此架構可縮短資料擷取時間、降低運算負載，並展現應用於即時光譜感測、邊緣人工智慧與可攜式智慧光電系統之潛力。

Traditional spectral recognition systems rely on external optical filters, gratings, or spectrometers. Although these systems provide good spectral resolution, their bulky size, high cost, and integration complexity limit their applications in portable and real-time sensing. To address this issue, this study develops a filter-free intelligent wavelength recognition framework by integrating deep learning with a solution-processed 4 × 4 p-Cu2O/n-Si self-powered photodetector array. The proposed framework exploits wavelength-dependent differences in transient photoresponses for wavelength classification. First, LSTM and BiLSTM models were employed to classify four wavelengths using truncated photocurrent signals. The results showed that the BiLSTM model with 64 hidden units achieved 100% classification accuracy using only the first 40 ms of the original 150 ms signal, demonstrating that early-stage transient signals contained sufficient discriminative information for wavelength recognition. Second, to improve model efficiency and interpretability, this study introduced a self-attention-enhanced MLP, enabling the model to automatically focus on highly discriminative temporal regions. The model achieved an accuracy of 95.73% using the full signal and still maintained an accuracy of 94.44% when using a shortened 20 ms dataset. These results indicate that the model can effectively extract the most discriminative temporal segments from the transient photoresponse, thereby reducing its dependence on complete measurement signals. Finally, to address the spectral aliasing problem associated with single-photocurrent-signal analysis, this study further proposes an Isc/Voc dual-channel recognition strategy to extract features related to carrier transport and charge accumulation. After integrating the dual-channel transient signals into an LSTM model, the system achieves a classification accuracy of 98.72% across the ultraviolet-to-near-infrared spectral range, effectively reducing recognition confusion between wavelengths with similar photocurrent responses. Overall, this study demonstrates that a p-Cu2O/n-Si self-powered photodetector array combined with truncated signal analysis, an attention mechanism, and Isc/Voc dual-channel fusion can realize fast, highly accurate, and physically interpretable filter-free wavelength recognition. This framework shortens data acquisition time and reduces computational load, showing strong potential for real-time spectral sensing, edge artificial intelligence, and portable intelligent optoelectronic systems.