Efficient terahertz absorption gas sensor with Gaussian process regression in time- and frequency-domain

Abstract

The terahertz region presents promising opportunities for quantitative gas sensing in practical environments thanks to its access to molecular fingerprints and immunity to particulate scattering. With Terahertz time-domain spectroscopy (THz-TDS), molecular absorption and dispersion spectra can be rendered using femtosecond-laser-based interferometric system. However, procurement of accurate spectroscopic parameters in this uncharted spectral region and quantitative reduction of the broadband data remain challenging. We propose and validate two end-to-end machine learning models based on Gaussian process regression (GPR) for efficient THz-TDS sensing in both time- and frequency-domain (TD and FD). Quantitative CO sensing is accomplished for proof-of-concept demonstration of the method. Both TD- and FD-GPR models demonstrate accurate predictions in the presence of measurement noise and ambient interference, and the root-mean-square errors are below 0.6 % with both simulated and experimental data. It is concluded that while the FD model benefits from immunity to dispersion distortions, the TD model noses out for its baseline-free nature and better elimination of temporally introduced noises. This work presents the first experimental demonstration of THz-TDS for quantitative gas sensing exploiting both TD and FD signals, showing promising prospects for real-time, multi-species, multi-parameter sensing under challenging scenarios such as chemically reactive or open-path atmospheric environments.