Yuichiro Kunai

Three-dimensional metastructures with nanoscale feature sizes exhibit unique properties compared with structures with larger feature sizes, but are difficult to fabricate. Here we introduce implosion carving (ImpCarv), a method for photopatterning vacancies of complex geometry throughout materials, followed by isotropic shrinkage (>10-fold). ImpCarv works by photoactivating sensitizers to generate reactive oxygen species that cleave a swollen hydrogel at defined points, followed by controlled shrinkage via dehydration. ImpCarv creates three-dimensional metastructures where the refractive index of each point throughout a material can be specified with nanoscale precision via material presence or absence. By leveraging refractive index programmability for precise phase control, we demonstrate an all-optical machine learning device with nanoscale neuron sizes operating at visible wavelengths. ImpCarv may thus support diverse applications in nanophotonics and nanotechnology.

Diffractive deep neural network (D²NN), a type of ONN that processes images using light propagation through free space, has been shown to be capable of performing various image processing tasks, yet it still relies partially on electrical signals. In this report, we achieved completely all-optical and continuous real-time processing of two-dimensional visible information directly using light from real objects without converting the input information into any electrical signals. Firstly, two image processing tasks, classification (MNIST) and detection of industrial parts, were performed as a verification of the capability of our novel D²NN optical information processor. Then, high-speed operation was investigated by two further tasks, classification of glass beads and detection of shapes printed on transparent sheets, in which the visible light from the sample was introduced to and processed by the processor without conversion to electrical signal. The results show that our implementation of a D²NN processor is capable of image processing on the nano-second order from the appearance of the sample in the observation area.

Optical processors, built with “optical neurons,” can perform large-scale high-dimensional linear operations at the speed of light. With the current advances in micro-fabrication, such optical processors can now be 3D-fabricated, but at limited precision, eventually leading to a model mismatch due to quantized optical weights. To address this issue, we propose a quantization-aware training framework. Our approach accounts for physical constraints during the training process, leading to robust designs. We numerically demonstrate that our approach can design state-of-the-art optical processors using diffractive networks for multiple tasks despite quantized learnable parameters. We thus lay the foundation upon which improved optical processors may be 3D-fabricated in the future.