Improved spectral filtering of broadband diffractive neural network by loss function engineering

Abstract

We engineer the loss function by removing the conventional physics-based energy constraint during the training of broadband diffractive neural networks (DNNs) to enhance their spectral filtering capabilities of supercontinuum light. Simulations show that compared to DNNs trained with conventional loss function, the suppression of out-of-band spectral intensities can be improved by three orders of magnitude, resulting in an extinction coefficient of 10−6. Additionally, the spectral resolution can be enhanced by over 50% with a 6.6% improvement of energy efficiency. These findings are corroborated by experiments conducted with a two-layer DNN. The proposed method holds promise for enhancing the performance of broadband DNNs across various applications, including spectral reconstruction, spectrum classification, and color image processing, among others.