Nanoimprinted Infrared Metasurface Multispectral Filters for Neural-Network-Assisted Low-Temperature Sensing

Low-temperature infrared sensing requires temperature information to be extracted from weak thermal radiation while preserving optical throughput. Nanoimprinted metasurface filters combine multispectral encoding with neural-network calibration to address this challenge. A single circular-ring geometry controls passband position, bandwidth, and overlap, with simulated tuning across 6.4-11.2 μm. Deep-ultraviolet stepper lithography and ultraviolet nanoimprinting transfer reusable master patterns to 1-inch zinc sulfide filters. Four broad, partially overlapping bands retain peak transmittance near 80% and provide complementary spectral inputs for blackbody measurements from -20°C to 20°C. Neural-network regression on the processed calibration curves achieves a mean absolute error of 0.160°C, compared with 0.434°C for one unfiltered input, a reduction of approximately 63%. Temperature-resolved comparisons also identify lower errors at the cold end of the range. By connecting geometric spectral design, replicable fabrication, and computational temperature retrieval, this approach offers a route to compact multispectral thermometers that make better use of weak infrared signals.