Development of a semi-supervised machine learning based noise filter for quantum cascade laser-coupled mid-infrared spectrometer

Abstract

The rapid and accurate quantitative determination of molecular rovibrational spectral features is highly desirable for applications in fundamental high-resolution spectroscopy, trace chemical sensing, and environmental monitoring. But significant spectral noise can impede the accuracy and precision of such spectral measurements as well as the precise classification of fine and hyperfine spectral fingerprints. Nevertheless, the denoising of weak rovibrational spectral data has long been a computational and experimental challenge. Here, we develop a new semi-supervised machine learning (SSML) denoising method combining unsupervised spectral eigenvector space dimensionality reduction of the spectral data matrix exploiting principal component analysis with supervised Fourier domain residual analysis for enhancing high-resolution rovibrational spectral signal quality. We discussed the detailed implementation of the SSML algorithm on the real experimental mid-infrared rovibrational spectral features of nitrogen dioxide (NO2) in the gas-phase acquired from a quantum cascade laser-coupled cavity ring-down spectroscopic method. Our results confirm the robustness and feasibility of the SSML approach and could lead to a wide range of pertinent applications in infrared spectroscopy including precise spectral signal analysis, concentration retrieval and performance enhancement of laser-coupled spectroscopic sensors.