Throughput-enhanced FTIR spectrometers with deep learning-based spectral recovery

Abstract

Fourier transform infrared (FTIR) spectrometers are widely employed for spectroscopic characterization with high optical throughput, owing to the circular Jacquinot Stop (J-Stop). While the throughput increases with the J-Stop diameter, it has been theoretically and experimentally found that the divergent beam entering the Michelson interferometer induces spectral broadening and shifting, resulting in the obvious degradation of spectral resolution and wavenumber accuracy. This work has validated the deep learning-based spectral recovery for numerically eliminating the spectral distortions in a rapid and accurate way, effectively replacing the previously proposed time-consuming iterative calculation. Considering the influence of spectral complexity on the spectral recovery, the spectra are segmented and classified in terms of peak number with the convolutional neural network (CNN), before the parallel spectral reconstruction with a set of generative adversarial networks (GANs). The spectral recovery has been performed on the transmission spectra of water vapor collected by a commercial FTIR spectrometer, demonstrating that the spectra of 4 mm-diameter J-Stop with ∼12 times higher throughput can be recovered to coincide well with those of 1 mm-diameter J-Stop. With the enhanced throughput, the signal-to-noise ratio of FTIR spectrometers can be improved by >6 times for weak-light detection.