Intelligent Inverse Design of Meta‐Absorbers Based on Reversible Deep Learning

Optical meta‐absorbers are artificially engineered materials capable of trapping light in a highly controllable manner, rendering them attractive for a wide range of applications. Artificial intelligence provides powerful tools for the efficient design of such complex structures. Among current methods, tandem neural networks (TNNs) have emerged as a mainstream solution for forward‐inverse modeling. However, TNNs require two separate networks, leading to high computational costs and limited accuracy—particularly in inverse design. In this work, we propose a novel and highly precise design paradigm based on reversible deep learning (RDL), which employs a single, bidirectionally trainable neural network. The RDL framework performs simultaneous training of the inverse and forward models within a unified architecture—addressing structural generation and spectral prediction, and significantly reducing computational costs. As a proof of concept, we demonstrate the design of near‐infrared meta‐absorbers using RDL, achieving an order‐of‐magnitude improvement in inverse prediction accuracy compared to TNNs, while using only 66.35% of the training parameters and requiring minimal memory and storage. This study introduces a flexible and efficient platform for the rapid and accurate design of meta‐absorbers tailored to specific absorption spectra, facilitating the intelligent development of advanced optical metamaterials and metadevices.