High‐Speed Design of Multiplexed Meta‐Optics Enabled by Physics‐Driven Self‐Supervised Network

Abstract The artificial intelligence (AI) can accelerate the meta‐optics design by rapidly predicting the transmission coefficients of individual meta‐atoms. However, extensive optimization iterations are usually required to complete the desired metasurface consisting of massive meta‐atoms. For designing meta‐holography, any change to the target image forces the whole process to repeat, resulting in lengthy computation time. Here, a physics‐driven self‐supervised network (PDSS‐Net) built upon AI‐assisted optimization frameworks are proposed to further expedite the design process. The encoder‐decoder module introduced into the PDSS‐Net can establish a mapping between the input holographic images and the output structural parameters of all meta‐atoms. After self‐supervised training, the network learns this mapping and enables iteration‐free inference for inputs beyond the training dataset. The design of 2K‐resolution, three‐wavelength‐multiplexed meta‐holograms is completed within one second, achieving a computational speedup exceeding 1000‐fold over conventional optimization‐based approaches. By retraining, more complex tasks are achieved as demonstrated in the design of both the wavelength‐polarization‐depth multiplexed scalar and vectorial meta‐holograms. This iteration‐free computational paradigm with adaptability in typical multiplexed meta‐optics can be applied to the intelligent design of multifunctional metasurfaces, facilitating large‐scale applications of meta‐devices.