Physics-guided neural network for channeled spectropolarimeter spectral reconstruction

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computational Imaging

Background:

• Traditional deep neural networks (DNNs) for channeled spectropolarimeter (CSP) reconstruction require extensive training datasets.
• This reliance on ground truth data limits their applicability and efficiency.

Purpose of the Study:

• To develop a novel reconstruction method for CSP that eliminates the need for training datasets.
• To leverage the physical model of CSP to guide the DNN optimization process for improved accuracy.

Main Methods:

• A physics-guided neural network (PGNN) approach was developed, integrating the complete physical model of CSP into a DNN.
• The method initializes DNN parameters randomly and uses gradient descent, constrained by the physical model, to estimate parameters and the mapping relationship.
• The physical model actively participates in the DNN parameter optimization, providing physical guidance.

Main Results:

• The PGNN method successfully reconstructs CSP data without requiring ground truth datasets.
• Simulations and experimental results demonstrate the superior performance and high-precision reconstruction capabilities of the PGNN.
• The physical guidance ensures that the DNN output aligns with physical principles.

Conclusions:

• The proposed physics-guided neural network (PGNN) offers a highly efficient and accurate method for CSP data reconstruction.
• This approach significantly reduces the dependency on large training datasets, promoting wider practical applications of CSP.