Robustness of parallel subnetwork-filtered diffractive deep neural networks

The diffractive deep neural networks (D²NNs) have been widely adopted due to their advantages of ultra-low latency, low power consumption, and highly parallel optical computing capability. However, its hardware implementation faces several challenges, among which the alignment of multiple diffractive planes remains a major obstacle to practical deployment. To address this issue, we propose a parallel subnetwork-filtered diffractive deep neural networks (PSF-D²NNs) architecture to enhance the robustness of conventional D²NNs against unavoidable multi-plane misalignment and other system errors during the experimental implementation. The robustness of the proposed network is enhanced through two strategies. Firstly, a dataset incorporating random misalignment errors of the modulation planes is introduced. Secondly, phase filtering factors are employed to smooth the modulation phase. The proposed network demonstrates strong performance enhancement with respect to misalignment both in simulations and experiments. In simulation, PSF-D²NNs achieve a misalignment tolerance up to 26 pixels with an MSE of 5.441 × 10^-5 and an SSIM of 0.7741, while the experimental tolerance exceeds 13 pixels with an MSE of 3.350 × 10^-6 and an SSIM of 0.9760.