Diffractive Optical Computing In Free Space

This article discusses the use of structured surfaces in free-space optical processors for deep learning applications. These processors use diffractive- and meta-units to process analog optical waves and perform statistical inference tasks. The trainable parameters of these processors are the transmittance coefficients of the diffractive unit cells, which are optimized using deep learning-based training on a digital computer. The resulting diffractive surfaces are then fabricated and assembled to form the physical processor.