Deep-learning-enhanced nonlinear holography in 3D nonlinear photonic crystal

Nonlinear holography in 3D nonlinear photonic crystals (NPCs) is capable of reconstructing multiple images at newly generated wavelengths, which has attracted increasing interest in optical storage and encryption. However, the storage capacity of nonlinear holography is generally limited by the crosstalk between different channels. In this work, we propose deep-learning-assisted high-capacity nonlinear holography to address this issue. By leveraging the exceptional recognition capability of the ResNet-18 convolutional neural network, the information at multiple channels can be well recognized (i.e., the accuracy reaches 88.1%) in the presence of severe crosstalk (i.e., a bandwidth overlap of 30% between neighboring channels). Importantly, the channel number is doubled in comparison to the previous reports. Our results provide a novel approach to bypass the negative effect of crosstalk for high-security and high-density optical information storage.