Machine learning-assisted ultra-selective tungsten metasurface absorber/emitter for high-efficiency Vis-NIR photothermal conversion

High-efficiency visible-to-near-infrared (Vis-NIR) photothermal conversion is essential for advancing solar thermophotovoltaic (STPV) systems, yet traditional designs are often limited by poor spectral selectivity and strict angular constraints. This paper introduces a machine learning-assisted design framework for an ultra-selective tungsten metasurface absorber/emitter that substantially suppresses angular and polarization sensitivity. A metal-insulator-metal (MIM) structure was designed, and a machine learning framework combined with finite-difference time-domain (FDTD) simulations were implemented to accelerate structure parameters optimization and predict absorption performance. The absorption performance of the structure was evaluated by comparing six different machine learning algorithms, among them, the XGBoost model demonstrating superior predictive accuracy, achieving R 2 score of 0.9804, mean absolute error (MAE) of 0.0043 and root mean square error (RMSE) of 0.0086. The proposed metasurface absorber achieves a remarkable 98.3% average absorptivity in the wavelength range 350–1000 nm, featuring a 99.8% absorption plateau in the range 450–700 nm. Crucially, the absorption performance exceeds 90% at wide incidence angles up to 60° under both transverse-electric (TE) and transverse-magnetic (TM) polarizations. Furthermore, the structure functions as an ultra-selective metasurface emitter, reaching a maximum spectral efficiency of 99.4% at 2000 K. These results demonstrate that combining machine learning with specialized MIM metasurfaces offers an accelerated pathway for designing high-performance thermal energy conversion systems, particularly for integration into silicon-based STPV cells.