Abstract

The speckle reconstructive spectrometer (RS) is revolutionizing the design paradigm of spectrometers, shifting from hardware-dominated architectures to algorithmically driven, computing-focused methodologies. However, traditional spectral reconstruction algorithms with pre-defined regularizes suffer from suboptimal adaptability, and similarly, deep learning methods struggle to handle unseen data types once deployed.

Here we propose a physics-aware spectral reconstruction framework named PhyspeNet, which consists of a convolutional neural network (CNN) and an empirical physics model, facilitating adaptive reconstruction of multiple spectral types without pre-training. The inherent structural priority in the CNN architecture provides an adaptive and remarkably powerful regularization, while the embedded empirical model endows our framework with generality across various dispersive devices, even those exhibiting complex light propagation beyond the scope of analytical models.

By leveraging PhyspeNet, we attain a resolving power of 7×10⁵, surpassing existing spectral reconstruction neural networks trained on polychromatic data. A 700 nm operating range in the near-infrared region is also realized, which, to the best of our knowledge, stands as the widest operational bandwidth ever reported in speckle spectrometry. We believe that this work constitutes a solid step toward adaptive speckle RSs and paves the way for advances in diverse fields, including high-dimensional light field detection, biomedicine and so on.