Optics×Artificial intelligence

Optics and AI are converging on two fronts: light as a physical substrate for neural computation (diffractive and photonic neural networks) and AI as the inverse-design engine for optics (metasurfaces, holography, computational imaging). As deep learning hits energy and latency walls, optical hardware promises massively parallel, low-power matrix operations — while neural nets crack optical design problems that defied analytic methods.

The bridge fields are unusually load-bearing: Holography, Photonics, and Nonlinear systems all already touch both communities, and 39 authors publish on both sides separately. The Adamic-Adar affinity (7.5) with 28 common neighbours signals a dense latent overlap despite zero direct co-publication yet. Field B's recent-share (0.047) is triple Field A's, meaning AI momentum is actively reaching toward optics rather than the reverse — a classic pre-collision asymmetry. Representative optics papers (metasurfaces, cryo-EM motion correction) are already computation-heavy, one MotionCor2 step away from full learned pipelines.

Groups fluent in both wave physics and modern ML training — photonics labs that have hired deep-learning talent, and AI-hardware teams recruiting optical engineers. The winners will own the co-design loop: differentiable optical simulators that let gradient descent flow through Maxwell's equations. Whoever controls the fabrication-to-training feedback path (metasurface foundry + autodiff design) captures the value.

A fully differentiable end-to-end optical-neural pipeline: a metasurface or diffractive optical layer whose physical parameters are trained jointly with a downstream digital network via autodiff through a Maxwell/wave solver, then fabricated and benchmarked on an imaging or classification task against an all-electronic baseline for accuracy-per-joule.