Phase Multiplexed Optical Computing: Reconfiguring a Multi‐Task Diffractive Optical Processor Using Illumination Phase Diversity

ABSTRACT

Optical computing based on diffractive networks attracted broad attention due to its low latency, low energy consumption, and parallelism. Given sufficient degrees‐of‐freedom, a diffractive network composed of spatially‐optimized surfaces can all‐optically implement an arbitrary complex‐valued linear transformation between an input and an output aperture. Here, we report a monochrome multi‐task diffractive network architecture that leverages illumination phase multiplexing to dynamically reconfigure its output function and implement a large set of complex‐valued linear transformations. Each of the T target transformations is encoded and addressed by a unique 2D illumination phase profile (“phase key”) applied to the input aperture, activating the corresponding transformation at the output field‐of‐view. A common diffractive network, jointly optimized with T phase keys, demultiplexes these encoded inputs and executes any of the T transformations. We show that a network with optimized diffractive features can accurately realize T distinct complex‐valued linear transformations for arbitrary complex input fields, where and refer to the input/output pixels, respectively. In a proof‐of‐concept numerical study, T = 512 transformations were implemented with negligible error. Compared with wavelength multiplexing, phase multiplexing yields lower transformation errors, enabling larger‐scale monochrome optical processors for computing and machine vision.