Rapid Wavelength Classification Using Deep Learning on Truncated Temporal Signals in Cu 2 O/Si Self-Powered Photodetector Arrays

Conventional machine learning approaches for wavelength recognition using photodetectors typically rely on complete time-series photocurrent curves to extract critical temporal features such as rise and decay times. However, acquiring the full response curve can be time-consuming and impractical for real-time or edge applications. In this study, we propose a deep learning framework utilizing LSTM and BiLSTM networks to classify four distinct wavelengths (365, 465, 560, and 730 nm) based on truncated photocurrent signals from a self-powered Cu 2 O/Si photodetector array. The BiLSTM model achieved perfect classification accuracy (100%) with only the first 40 ms of the 150 ms signal when configured with 64 hidden units, demonstrating the feasibility of early-stage inference. In contrast, LSTM required the full temporal profile for comparable performance. The BiLSTM also exhibited strong robustness across varying training/test splits and random initializations, highlighting its generalization and reproducibility. These results demonstrate the potential of combining truncated temporal data with bidirectional deep learning to enable fast, efficient, and filter-free spectral sensing suitable for real-time edge-deployed applications.