Femtojoule optical nonlinearity for deep learning with incoherent illumination

• DOI: 10.1126/sciadv.ads4224
• Publication Date: 2025-01-31T18:58:41Z
• Authors: Qixin Feng; Can B. Uzundal; Ruihan Guo; Collin Sanborn; Ruishi Qi; Jingxu Xie; Jianing Zhang; Junqiao Wu; Feng Wang
• Keywords: Photodiode; Optical power; Optical computing

Optical neural networks (ONNs) are a promising computational alternative for deep learning due to their inherent massive parallelism linear operations. However, the development of energy-efficient and highly parallel optical nonlinearities, critical component in ONNs, remains an outstanding challenge. Here, we introduce nonlinear microdevice array (NOMA) compatible with incoherent illumination by integrating liquid crystal cell silicon photodiodes at single-pixel level. We fabricate NOMA more than half million pixels, each functioning as analog rectified unit ultralow switching energy down 100 femtojoules per pixel. With NOMA, demonstrate multilayer network. Our work holds promise large-scale low-power computer vision, real-time image processing.