Interpretability in near-infrared (NIR) spectroscopy: Current pathways to the long-standing challenge

Abstract

Near-infrared (NIR) spectroscopy holds a unique position in analytical chemistry, contrasting excellent utility with intrinsic spectral complexity. Challenges arise from high dimensionality and limited interpretability of its spectral data. Applied NIR analysis evolved to rely on multivariate data analysis, with various statistical parameters providing information sufficient to control and maintain the calibration process, ultimately delivering reliable predictions. However, in NIR spectroscopic analysis the key analytical insights derive from well-defined molecular vibrations, a factor often not fully explored in the routine. Recent advancements have tackled many of these persistent issues through refined chemometric approaches, theoretical simulations of NIR absorption bands, or novel approaches to deal with physical footprint of the sample. At the same time, trending data science approaches such as deep chemometrics introduce own challenges in model transparency. These developments dynamically alter the outlook on interpretability in NIR spectroscopy; the review critically examines the emerging potential at their intersection.