Performance Benchmarking of Machine Learning Models for Terahertz Metamaterial Absorber Prediction

Abstract

This study presents a polarization-insensitive ultra-broadband terahertz metamaterial absorber based on vanadium dioxide (VO2) and evaluates machine learning methods for predicting its absorption performance. The structure consists of a VO2 metasurface, a MF2 dielectric spacer, and a gold ground plane. It achieves more than 90% absorption between 5.72 and 11.11 THz, covering a 5.38 THz bandwidth with an average absorptance of 98.15%. A dataset of 9,018 samples was generated from full-wave simulations by varying patch width, dielectric thickness, and frequency. Six regression models were trained: Linear Regression, Support Vector Regression, Decision Tree, Random Forest, XGBoost, and Bagging. Performance was measured using adjusted R2, MAE, MSE, and RMSE. Ensemble models achieved the best results, with Bagging reaching an adjusted R2 of 0.9985 and RMSE of 0.0146. The workflow offers a faster alternative to exhaustive simulations and can be applied to other metamaterial designs, enabling efficient evaluation and optimization.