Compact photonic neurons using periodic nanohole arrays in metasurfaces

The increasing demand for computational power is beginning to exceed the limits of traditional electronic architectures, which are constrained by energy efficiency and scalability. Optical neural networks present a promising alternative that employs the speed and parallelism of light to conduct complex computations while using minimal power. Although diffractive optical neural networks (DONNs) enabled by metasurfaces have shown some promise, conventional metaline-based neurons in 1D/2D metasurfaces require large footprints, restricting their high-density integration. This study introduces the use of periodic nanohole arrays (PNAs) in metasurfaces to provide a compact and precise optical neuron structure. We found that PNA-based neurons demonstrated enhanced phase modulation in our simulations, and our experimental results verified their performance within single- and three-layer DONNs, achieving classification accuracies of 88.9% and 82.2% on the Iris dataset with neuron densities of 200 in a 0.075-mm2 footprint and 600 in a 0.15-mm2 footprint, respectively. These findings demonstrate that PNA structures offer compact and scalable solutions that can improve the performance of on-chip optical computing systems.