Machine learning-accelerated sustainable inverse design of ultra-wideband THz metamaterial absorber

Terahertz absorbers are progressively relevant to sustainable energy and sensing where material and computational efficiency are essential. We demonstrated a six-layer IGIGIM (Insulator–Graphene–Insulator–Graphene–Insulator–Metal) metamaterial absorber that achieves >90% absorption across 12.95 THz (6.07–19.02 THz), peaking at 99.99% near 11.85 THz. The structure maintains its performance up to 60° for both polarizations, demonstrating good tolerance to incidence-angle variations. Rather than relying on full-wave simulations, we developed a machine learning surrogate trained on FDTD (Finite-Difference Time-Domain) data to accelerate design exploration. Bayesian optimization identified high-performing designs using ∼60% fewer simulations than random search. Multi-objective analysis further revealed viable trade-offs among bandwidth, total thickness below 30μm, and angular robustness, supporting compact designs with reduced computational and material requirements. SHAP analysis showed that incidence angle dominates performance variation. Uncertainty quantification further allowed us to flag low-confidence predictions for selective FDTD validation, creating a reliability-aware workflow while reducing unnecessary computational effort.