Machine learning-based prediction of spatially and frequency-selective optical field coupled absorption in random metasurfaces

Abstract

Achieving frequency and spatially selective optical field coupling represents a crucial avenue for simplifying metasurface architectures. Given the intricacy of promptly analyzing the optical properties of micro- and nano-devices with numerous structural parameters, this study develops a high-performance prediction terahertz (THz) metasurface model based on machine learning. A core objective of the model is to qualitatively assess the significance and impact direction of each parameter. The investigation systematically evaluates two models (uniform and linear) comprising randomly distributed micrometer resonance spheres. The optimized metasurfaces exhibit remarkable performance, achieving average absorptions of up to 86.12% and 87.82% across the 1–10 THz band range. The resonant fields are distributed in distinct spatial regions at different frequencies, showcasing selectivity in coupling spatial and frequency optical fields. The interpretable analysis derived from machine learning aligns closely with physical mechanism investigations based on electric field distributions, establishing a high degree of consistency. This convergence opens innovative perspectives for designing high-performance THz wave micro- and nano-devices.