Quantized THz Diffractive Optics Design via Automatic Differentiation

Diffractive optical elements (DOEs) in the THz range provide unparalleled, multifunctional control over radiation. Neural network-based design approaches offer significant potential in optimizing the unique phase maps of DOEs, enabling nearly arbitrary modulation of radiation. However, these neural network (NN) design methods typically require continuous values for the DOE phase profile synthesis. Often, the methods do not consider the physical quantization in DOE fabrication due to the discrete material layers produced with 3D printers. To address this, we apply automatic differentiation, a technique commonly used in neural networks, to develop a novel method for optimizing the phase profile of heavily quantized DOEs. Our simulation experiments show that this approach facilitates fast and flexible DOE design, while considering the fabrication limitations.