Dynamic Long‐Wave Infrared Spectral Imaging With Bayesian‐Optimized Dual‐Phase‐Change‐Material Metasurface Encoder

ABSTRACT Computational spectral imaging in the long‐wave infrared (LWIR) band offers a compact solution for thermal sensing but remains limited by acquisition speed, system complexity, and cost, hindering their practical deployment in compact and reconfigurable systems. In this paper, a dynamic LWIR computational spectral imaging system enabled by a Bayesian‐optimized dual‐phase‐change‐material (dual‐PCM) metasurface spectral encoder and a Transformer‐based reconstruction network is demonstrated. The encoder employs vertically stacked Ge 2 Sb 2 Se 4 Te 1 (GSST) and Sb 2 S 3 metasurface pillars with independently tunable crystalline fractions, enabling multi‐level, reconfigurable spectral encoding with high modulation contrast and low optical loss across the LWIR band. A Bayesian optimization‐based inverse design framework is developed to efficiently minimize spectral correlation among encoding functions over multiple crystallinity states, reducing the correlation from 0.9918 to 0.1668 for the optimized encoder. With the optimized spectral encoder and the reconstruction network, the system achieves an effective system‐level spectral resolution of 100 nm, with a spectral fidelity exceeding 99.81% and a PSNR of 46.50 dB. This work provides a compact, intelligent, and scalable solution for high‐performance LWIR spectral imaging, paving the way toward miniaturized and programmable thermal spectral sensing systems.