Terahertz Diffractive Deep Neural Networks via Polarization‐Conversion Metasurfaces for Digit Recognition

ABSTRACT

To overcome the high energy consumption of conventional computing and the integration barriers of bulky diffractive deep neural networks (NNs), this study proposes a compact NN architecture using terahertz (THz) metasurfaces. Specifically, meta‐atoms are engineered to achieve a high polarization conversion transmission coefficient (exceeding 0.9) across the 0.55–0.65 THz range while providing full phase modulation. Based on these meta‐atoms, a THz optical neural network featuring three cascaded hidden layers is constructed. Simulations demonstrate that patterned electromagnetic waves, after being modulated by the three‐layer structure, can be converged onto designated detection regions. The system successfully recognizes a nine‐class subset of the MNIST data set comprising digits 0–6, 8, and 9 at 0.6 THz. This work demonstrates the viability of using THz metasurfaces for compact optical AI hardware with a passive metasurface inference core and next‐generation intelligent sensing.