AI-assisted ultrafast and efficient terahertz spectrometer using frequency selective surfaces

We introduce a novel method for ultrafast selective multispectral terahertz (THz) spectroscopy, combining broadband THz pulses, Frequency Selective Surfaces (FSS), and a Schottky diode energy sensor. Traditional THz spectroscopy is costly time-consuming and expert-operated. Our system answers to these challenges by not requiring to obtaining the time trace of the electric field of the THz signal thus essentially simplifying the system. Our system efficiently identifies samples by analyzing distinct spectral signatures. Experimental results demonstrate the method's ability to distinguish samples with similar THz absorption coefficients and refractive indices, even without clear fingerprint features. Validation on paper samples with closely matched THz properties confirmed successful differentiation through data averaging and normalization. We also applied k-fold cross-validation with a neural network for multi-class classification, achieving a training accuracy of 94.5% and an average testing accuracy of 94%. This approach offers robust real-time spectroscopic identification and potential for industrial applications and predictive modelling of THz signals.