Reinforcement Learning-Optimized Freeform Metasurface for Light Output Enhancement of Scintillators

Conventional photonic nanostructures with fixed shapes are capable of enhancing the light output of scintillators based on the partial elimination of internal total reflection. However, the enhancement effects are limited due to the relatively narrow design space. In this work, a two-dimensional freeform metasurface is proposed to achieve a light output enhancement effect exceeding the conventional fixed-shaped photonic structures. The freeform metasurface is designed by the exploration of a deep Q-network agent in reinforcement learning. As a proof-of-concept, the widely used LYSO scintillator is chosen. Simulated results show that the LYSO coating with an optimal freeform metasurface yields a maximum light output enhancement of 4.7-fold and an average enhancement of 2.1-fold over the full emission angle, exceeding the previously reported enhancements from nanostructures with fixed-shape. The method efficiently explores the vast configuration space by searching only approximately 10,000 of 10¹⁷ configurations and surpasses the feasibility of exhaustive search while overcoming limitations of genetic algorithms and adjoint-based optimization in high-dimensional design, thereby establishing a useful paradigm for designing high-performance scintillator metasurfaces.