PSO-Based Optimization of Material Parameters for Diffractive Deep Neural Networks

Optical computing is increasingly regarded as a promising alternative to conventional electronic computing due to its inherent advantages in speed and energy efficiency. Diffractive deep neural networks (D²NNs), as a typical optical computing architecture, suffer from insufficient feature extraction capability and poor flexibility in nonlinear classification. To tackle these issues, a hybrid D²NN architecture is proposed. In this hybrid D²NN architecture, feature extraction is accomplished by cascaded optical diffraction layers, whose material parameters (relative permittivity ε and permeability µ) are optimized using particle swarm optimization (PSO) to enhance light field modulation. Specifically, three diffractive layers are retained and a linear gradient distribution of ε and µ is applied to simplify the fabrication process, thereby enabling the refractive index n to be modified for improved feature modulation. Next, nonlinear classification is facilitated by fully connected electronic layers appended after the diffractive stages, which compensate for the linear constraints inherent to purely optical systems. Consequently, the proposed hybrid architecture combines efficient optical feature extraction with flexible electronic classification. Experimental results on the MNIST, Fashion-MNIST and grayscale CIFAR-10 datasets demonstrated that classification accuracies of 96.8%, 89.73% and 53.32% were achieved by the hybrid architecture of D²NN, respectively, which significantly outperforms the original D²NN.