Electrochemically‐Reconfigurable Diffractive Neural Network With Metasurfaces

Authors:

Abstract
Optical neural networks (ONNs), particularly diffractive neural networks (DNNs), offer a promising computational paradigm owing to their high speed, intrinsic parallelism, and low energy consumption. However, most demonstrated DNNs are static, with their functionalities fixed after fabrication. Although several reconfigurable architectures such as spatial light modulators (SLMs) and electrically controlled digital‐coding metasurfaces have been demonstrated, they often suffer from high power consumption and limited scalability. Here, we introduce an electrochemically reconfigurable diffractive neural network (ECDNN) based on a PANI‐functionalized Pancharatnam‐Berry (PB) phase metasurface, whose optical response is reversibly controlled through electrochemical redox reactions. Using a simultaneous multi‐objective optimization (SMOO) strategy, different computational tasks can be mapped to distinct state configurations. As a proof of concept, we demonstrate a single‐layer ECDNN capable of performing two classification tasks. The system exhibits response times of 30–40 ms for optical transmission under sub‐volt driving voltages, enabling low‐power operation and facilitating compatibility with low‐voltage driving electronics. Numerical studies further extend the state‐conditioned framework to a two‐layer architecture with four predefined state combinations. This work provides a fast, energy‐efficient, and high‐neuron‐density hardware platform for brain‐inspired optical computing.

This publication has 34 references indexed in Scilit: