Machine Learning-Driven Ultra-Broadband Terahertz Multilayer Metamaterial

The design of subwavelength-scale metamaterials is challenging due to their intricate electromagnetic properties, making traditional simulation approaches computationally intensive and time-consuming. In this research, we introduce a machine learning (ML)-based method for the optimal design of an ultra-broadband, multi-stacked terahertz metamaterial. The proposed metamaterial features a cross-shaped resonator atop hexagonal resonators, separated by a polyimide spacer, to achieve ultra-wideband THz transmission. We developed forward and inverse ML models using Ridge Regression (RR), Support Vector Regression (SVR), Random Forest Regressor (RF), Neural Network (NN), and XGBoost Regressor to predict transmission spectra from structural parameters and vice versa. Random Forest and XGBoost, which had the lowest average mean absolute errors, were identified as the most effective models for achieving the desired ultra-broadband spectrum. The Random Forest model was further validated by predicting the structural parameters of a band-reject filter with an ultra-broadband bandwidth (FWHM > 250 GHz) that was not part of the training data. These predictions were confirmed through simulations. We then fabricated the ML-generated design and experimentally analyzed it using THz time-domain spectroscopy (THz-TDS). The measured transmission amplitudes closely matched the predictions from the ML models, further validating the accuracy of the ML predictions. To gain insight into the broadband transmission mechanism, we analyzed the electric field distribution at various resonance frequencies and investigated the near-field coupling between resonators using a hybridization mode-splitting model.