Enabling Reinforcement Learning to Two-Dimensional Freeform Metasurface Inverse Design via Fourier Level-Set Representation
Abstract
Freeform metasurfaces offer a vast design space to achieve performance beyond the conventional fixed-shape metasurfaces, but navigating this space efficiently remains a significant challenge. Reinforcement learning (RL), as one of the modern inverse design methods, holds the potential to handle the immensity of the design space without requiring prior training data. Yet, most demonstrations in freeform metasurface design have been limited to one-dimensional (1D) grating design problems and only a limited number of studies has been applied to two-dimensional (2D) metasurface. The exponential scaling of the design space and the expensive three-dimensional electromagnetic simulations make 2D metasurface inverse design considerably more challenging. To overcome this limitation, we propose Fourier level-set reinforcement learning (FLRL), which integrates the Fourier level-set representation and reinforcement learning into a unified optimization framework. The Fourier level-set representation effectively shrinks the design space by producing smooth, connected structures, thereby eliminating unnecessary designs. By leveraging RL’s exploration capabilities within this reduced design space, FLRL successfully optimizes high-performance 2D metasurfaces, exceeding previous records achieved with grid-cell representation and achieving higher efficiency compared to gradient-based method on the same level-set representation. This work demonstrates the potential of level-set representation to enable the adoption of machine learning methods for highly complex 2D metasurface inverse design.