All-Optical Dual-Angle Diffractive Neural Networks for 2D Direction-of-Arrival Estimation

Accurate direction-of-arrival (DoA) estimation in real time is essential for stable terahertz (THz) communications. To overcome the limitations of previous diffractive methods that are restricted to 1D sensing or depend on rotation, and to advance their practical deployment, we propose and experimentally validate an all-optical dual-angle diffractive neural network (DA-DNN) capable of simultaneous 2D angular classification. At the heart of our design is a diffractive aperture placed before the neural network as a physical encoder. It converts the phase-only angular information carried by the incident wave into a complex optical field enriched with amplitude features, thereby producing robust spatial encoding sensitive to angle. The system is modeled using the Rayleigh-Sommerfeld diffraction integral and optimized through end-to-end backpropagation. We fabricate a compact, monolithically integrated DA-DNN device using 3D printing at high resolution and evaluate it through full-wave simulations and THz experiments that emulate communication environments. The system accurately recognizes 64 angular combinations over a $\left(\left(-40°,40°\right)\right)$ range, providing a passive and scalable optical solution with low latency for THz communications.