Multitask Diffractive Deep Neural Networks Enabled by Joint Distance-Frequency Encoding

Diffractive deep neural networks (D²NNs) offer a promising route toward high-speed wave-based computing, but most existing architectures are restricted to a single fixed function, whereas multitask implementations often require additional hardware complexity. Here, a multitask D²NN architecture based on joint distance and frequency encoding is proposed and demonstrated. The interlayer distance is introduced as a physically reconfigurable task-encoding dimension and is combined with a dual-frequency phase-independent metasurface to enable task multiplexing within a single passive hardware system. Scalar diffraction simulations on the corresponding testing datasets yielded accuracies of 99.7% for 9-class orbital angular momentum (OAM) classification, 86.4% for 9-class direction recognition, 98.8% for 4-category programmable logic-gate operations, and 95.4% for 4-class MNIST digit recognition (Digits 0-3). Full-wave electromagnetic simulations further support the physical feasibility of the designed metasurface implementation, whereas microwave measurements on selected representative inputs provide proof-of-concept evidence of task-dependent energy routing and distance-enabled task switching. This joint-encoding strategy provides a compact and resource-efficient route for multifunctional intelligent wave-based computing systems.