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

Diffraction Neural Networks (DNNs) is a novel optical computing architecture that combines wave optics with deep learning methods for high-speed parallel information processing. We propose a novel loss function design that eliminates traditional physics-based energy constraints to enhance its spectral filtering capability for supercontinuum light. Simulations show that compared with DNNs trained with traditional loss functions, the suppression of out-of-band spectral intensity can be improved by three orders of magnitude, with an extinction coefficient of 106. In addition, spectral resolution can be improved by over 50%, and energy efficiency can be increased by 6.6%. Experimentally, we designed an experimental setup consisting of two phase modulation layers based on a single spatial light modulator and a mirror facing it. The power efficiency of this design is more than 16 times higher than that of cascaded structures using beam splitters. It has two working modes: as a two-layer DNN or as a one-layer DNN, with the other layer serving as an information input layer. We have demonstrated the superior performance of the new loss function through experiments conducted in the two-layer DNN working mode of the experimental setup, and the experimental results are highly consistent with the simulation. The proposed method is expected to improve the performance of broadband DNNs in various applications, include Diffraction Neural Networks (DNNs) is a novel optical computing architecture that combines wave optics with deep learning methods for high-speed parallel information processing. We propose a novel loss function design that eliminates traditional physics-based energy constraints to enhance its spectral filtering capability for supercontinuum light. Simulations show that compared with DNNs trained with traditional loss functions, the suppression of out-of-band spectral intensity can be improved by three orders of magnitude, with an extinction coefficient of 106. In addition, spectral resolution can be improved by over 50%, and energy efficiency can be increased by 6.6%. We have demonstrated the superior performance of the new loss function through experiments conducted, and the experimental results are highly consistent with the simulation. The proposed method is expected to improve the performance of broadband DNNs in various applications, including spectral reconstruction, spectral classification, and color image processing.ing spectral reconstruction, spectral classification, and color image processing.