120-Channel High-Fidelity Optical Multiplexing Using a Minimal Meta-Atom Set

High-capacity optical multiplexing is crucial for next-generation intelligent information systems, spanning high-speed communications, optical computing, and immersive displays. A fundamental challenge is to maximize the number of independent channels while maintaining high fidelity, compact footprint, and low structural complexity, constraints that severely limit existing diffractive optical elements. Here, we overcome these long-standing trade-offs by introducing a multi-dimensional reconfigurable diffractive neural network (MRDNN) that enables high-capacity, high-fidelity optical multiplexing within an exceptionally compact and simple architecture through the co-design of predictive neural optimization and multi-physical-dimension multiplexing. Experimentally, a dual-layer MRDNN realizes 24 independent channels for distinct functionalities, including focusing, beam steering, and holographic image reconstruction. Owing to its excellent scalability, only four metasurface layers comprising a minimal library of 200² meta-atoms achieve 120 distinct optical channels with low crosstalk and high fidelity (measured mean structural similarity index measure of over 0.8), representing a record experimental multiplexing capacity with one of the smallest normalized footprints and simplest structural configurations in a metasurface-enabled system. This work establishes a scalable and efficient paradigm for extreme optical multiplexing in multifunctional and reconfigurable photonics systems.