Diffractive Optics

last updated 2026-08-31 · +1 sources in last 30d

Physics / mechanism

Diffractive optics shape light by imposing a spatially varying phase (or amplitude) profile on a wavefront and letting free-space propagation convert that modulation into a desired field distribution. Unlike a refractive lens, which accumulates phase through bulk material thickness, a diffractive element encodes the phase in surface relief or sub-wavelength structure, so the functional layer can be thin and lithographically or additively fabricated. Cascading several such layers, separated by propagation distances, gives a multi-layer transform whose input-output mapping is set entirely by the fixed phase patterns.

The diffractive deep neural network (D²NN) formulation treats each layer as a trainable set of phase pixels: the layers are optimised in simulation by deep learning, then fabricated as passive elements (3-D printed or lithographic), after which light diffracting through the stack performs the computation at the speed of light with no power consumed beyond the illumination. The design parameters are therefore layer count, pixel pitch relative to wavelength, inter-layer spacing, and the achievable phase depth and fabrication fidelity of each pixel.

The same multi-layer phase-modulation architecture can be trained for routing rather than classification. A D²NN-style stack has been used to implement arbitrary simultaneous unicast and multicast connectivity in an optical switch, removing the lossy power splitters that limit multicast optical circuit switching scalability, and to add wavelength selectivity so that the same hardware routes in both space and wavelength.

Efficiency is the principal physical constraint at short wavelengths. In full-field transmission X-ray microscopy, the reduced efficiency of diffractive optics at high photon energies limits performance, alongside the difficulty of matching the numerical aperture of the illumination to that of the objective; free-form diamond refractive optics were used to address the illumination-efficiency bottleneck, with full-field nano-imaging demonstrated at 20 keV.

Competitive landscape

ApproachWhere the sources place it
Multi-layer diffractive stacks (D²NN)Passive, power-free optical computation and machine-learning inference at the speed of light; also arbitrary unicast/multicast routing with wavelength selectivity
Splitter-based multicast optical circuit switchingScalability capped by splitter loss, which the diffractive stack eliminates
Free-form refractive optics (diamond)Chosen over diffractive optics for high-energy X-ray illumination, where diffractive efficiency falls off

The sources therefore show diffractive optics winning on passivity and on the ability to encode arbitrary trained transforms, and losing on raw efficiency in the hard X-ray regime, where refractive freeform elements were preferred.

Evidence base

Frontier (open questions)

Synthesised 2026-08-31 from 3 KB sources by the resynth pipeline; citations are KB source slugs.

Frontier questions