Ultracompact Imaging Devices by Deep Diffractive Neural Networks

Deep diffractive neural networks (D2NNs) that combine optical diffraction physics and deep learning capabilities have recently been utilized for the inverse design of multi-layered diffractive devices that include pulse shaping elements, broadband filters, and ultracompact dual band lenses with designed spectral and spatial responses. D2NNs are directly trained using error backpropagation within the diffractive layers without the need of training datasets. Moreover, when combined with adaptive loss weight algorithms, these networks achieve a more efficient inverse design of complex phase devices compared with purely data-driven neural network approaches.<br/>In this talk, we present our work on the design, development, and experimental characterization of novel ultracompact diffractive devices based on adaptive-weight D2NN (a-D2NN). Specifically, based on the highly transparent hydrogenated amorphous silicon platform, we will introduce two-layer diffractive devices that operate over multiple visible and infrared spectral bands with designed point-spread functions (PSFs) and optimal focusing efficiency beyond the theoretical limit of single-lens devices. Moreover, we will discuss ultracompact focusing spectrometers (100µm-footprint) that achieve &lt;10nm spectral resolution and &gt;50% diffraction efficiency across the visible spectrum. Finally, we will introduce a-D2NN for the inverse design of hyperuniform phase plates (HPPs) and experimentally demonstrate high-fidelity image reconstruction based on fabricated 4-level diffractive devices in silicon. The flexible a-D2NN approach discussed here enables the engineering of multi-layer diffractive optical devices with targeted spectral and spatial responses for advanced multiband imaging and microscopy applications.