Augmentations for selective multi-species quantification from infrared spectroscopic data

Sensitivity and selectivity are arguably the two most important qualities in a new sensor design. While many spectroscopic sensors developed in laboratory conditions achieve high sensitivity and selectivity, they are not always applicable to real-world conditions. Challenges in real-world applications come from corruptions like noise and interference. This study leverages machine learning methods for accurate and robust quantification under such corruptions. We propose simple yet effective augmentation strategies that promote robustness against unknown interference. The performance of the proposed augmentations is compared under varying levels of interference and noise. We demonstrate our methodology for a gas sensing application using infrared spectroscopy data. We focus on quantifying common volatile organic compounds (VOCs) in a realistic scenario with several unknown interfering species. The findings of this work put us a step closer to creating a robust and widely-applicable sensing platform. Synopsis: Interference is a major obstacle in the development of reliable spectroscopic gas sensors. This study explores data augmentation strategies to tackle unknown interference and noise to propose a sensing strategy for volatile organic compounds. [Display omitted] •Propose a methodology for real-world sensing applications.•Compare augmentations for unknown interference.•Quantify performance against various levels of noise and interference.•Identify Tanimoto similarity of the Morgan fingerprint as a potential screening method.