Optimizing data analysis for broadband mid-infrared absorption spectroscopy: A hybrid dataset approach.

BACKGROUND

Broadband mid-infrared spectroscopy for gas sensing is a rapidly progressing field with a main frontier in novel mid-infrared ultra-broadband laser sources. These sources provide increasingly complex and high-resolution spectra that are useful for highly sensitive multispecies (trace) gas detection. However, laser-based sources also add a high level of instrument-specific noise and baseline drifts. To efficiently handle the complexity of data and simultaneously overcome the noise and baseline drift, a demand arises to assess and enhance the processing of the acquired spectra.

RESULTS

We present a simulation-based approach that improves the detection of gas compounds from broadband mid-infrared spectra. The central idea is to construct a realistic simulation environment in which the data processing of any model of choice can be improved. This can benefit both commonly used processing techniques, such as classical least squares (CLS) fitting, which can be fine-tuned, and statistical models such as partial least squares (PLS), which demand a (relatively) large realistic training set. In addition, we can estimate an instrument- and application-specific detection limit a priori. The simulated spectra are made by combining simulated absorbance spectra with measured (featureless) background intensity spectra, which is crucial to incorporate instrument-specific effects such as baseline drifts. The resulting hybrid dataset is scalable in size and complexity. The workflow was applied to real-life measurements in the 8-11 μm wavelength region to detect trace levels of acetone in CO- and water vapor-rich exhaled human breath samples. Both CLS and PLS models could be considerably improved by using our proposed approach.

SIGNIFICANCE

The workflow presented here provides a means to optimize, train, and assess data-processing techniques for broadband mid-infrared gas spectra. The approach is widely applicable: it can be implemented for any gas absorption spectroscopic instrument with broadband coverage, as long as it is possible to determine the instrument-specific characteristics, and it can improve and evaluate a wide variety of data-processing techniques.