Translation Multiplexing of Optical Diffraction Neural Network

ABSTRACT Optical neural networks (ONNs) face critical scalability barriers due to the manufacturing complexity of large‐area metasurfaces and static multiplexing paradigms. Here, we introduce a translation multiplexing optical neural network (TMONN) framework that achieves dynamic lateral‐shifting multiplexing—distinct from polarization or angular momentum‐based approaches—through deep learning‐optimized remapping of redundant spatial information. By encoding overlapping data streams into programmable DMD‐SLM modulation layers and integrating a closed‐loop self‐calibration system for real‐time aberration correction, TMONN reduces hardware footprint while preserving computational resolution. Our architecture demonstrates more than 500% efficiency improvements over conventional ONNs and maintains robust performance (< 0.16 MSE degradation at OR = 7/8 multiplexing) across time‐varying tasks, achieving SSIM > 0.7 and PSNR > 11 dB for temporal topological sequences and sub‐0.016 MSE in medical CT slice reconstruction. The resolution‐preserving multiplexing, validated through 9‐frame parallel processing without quality loss, bridges computational optics with adaptive deep learning, offering a scalable pathway toward energy‐efficient optical computing platforms for dynamic real‐world applications.