Deep-Learning-Assisted Inverse Design of Pixelated Ultra-Wideband Terahertz Metasurface Absorbers With Refractive-Index Sensing Capability
Abstract
Terahertz (THz) metasurface absorbers have attracted significant interest for wideband energy harvesting, electromagnetic stealth, and sensing applications. However, the design of ultra-wideband THz absorbers still relies heavily on empirical optimization and time-consuming full-wave simulations, particularly when high-dimensional pixelated topologies are involved. Here, we propose an autonomous inverse-design framework for pixelated THz metasurface absorbers, in which a deep-learning predictor and grey wolf optimizer (GWO) are combined to efficiently search a high-dimensional binary topology space. The optimized absorber exhibits an average absorption of 96.4% over the 2.17–10.32 THz frequency range, corresponding to a relative bandwidth exceeding 130%. Furthermore, the absorber maintains absorptivity above 70% under incident angles up to 70°, demonstrating excellent angular stability. The ultra-wideband absorption arises from the combined effect of four resonances, and the multipolar decomposition analysis reveals that these resonances are primarily dominated by electric-dipole responses. In addition, the high-frequency resonance exhibits refractive-index sensing capability with a sensitivity of up to 2.546 THz/RIU. The proposed topology-driven inverse design framework provides an efficient and scalable route for the automatic development of high-performance metasurface devices across microwave, THz, and optical frequency regimes.