Deep Neural Network-Based Optical Parameter Extraction and Material Classification using Terahertz Time Domain Spectroscopy

Terahertz time-domain spectroscopy has emerged as an effective technique for extracting optical properties from materials and subsequently determining their identity. Traditional optical extraction algorithms rely on analytical formulas or numerical iterative algorithms, and encounter limitations when faced with variables such as unknown material thicknesses, experimental misalignment, and the Fabry-Perot effect. In this study, we propose a novel approach leveraging recurrent neural networks, specifically the gated-recurrent unit (GRU), for time-series prediction. Our approach utilizes GRU to accomplish two primary objectives: (i) automating the extraction of optical parameters and (ii) classifying materials without prior knowledge of their optical properties. Experimental validation is conducted using materials with various thicknesses, including Endur™RGD450 and polydimethylsiloxane, demonstrating the accuracy and reliability of the proposed GRU model in automating the extraction of optical material properties and enabling materials classification.