Machine learning driven design and optimization of broadband metamaterial absorber for terahertz applications

Abstract This article presents a novel broadband metamaterial absorber (MMA) that leverages machine learning to enhance performance in terahertz applications. The MMA consists of a vanadium dioxide (VO2) resonator placed above a dielectric spacer and a conductive ground layer. By altering the conductivity of VO2, the MMA modulates its absorption characteristics. At high temperatures, the absorber demonstrates over 80% absorption across the frequency range of 1.22 to 3.14 THz, achieving peak absorption levels of 99.99% at 4.84 THz and 99.90% at 5.82 THz. The study further explores how geometric adjustments influence the absorption characteristics. The MMA exhibits polarization independence and wide-angle absorption, further enhancing its adaptability. Machine learning models—k-nearest neighbor (KNN), LightGBM (LGBM), and Histogram-based Gradient Boosting (HGB)—are employed to predict absorption coefficients at intermediate frequency, periodic dimensions, and substrate thickness. These regression models are evaluated using root mean squared error (RMSE), mean absolute error (MAE), and R 2 score across test sizes ranging from 0.4 to 0.6. The KNN model demonstrates outstanding performance, with R 2 values of 0.9977 for feature ‘a’ and 0.9982 for feature ‘ h s ’ at a test size of 0.4. The regression analysis suggests that ML-based MMAs can reduce simulation time and resource usage by up to 60%. This work establishes a framework for broadband MMA design using machine learning techniques, offering promising applications in terahertz modulation, sensing, and tunable devices.