Spectrometer Steers
HITRAN simulations of water and methane are used as inputs to an artificial neural network (ANN) that provides water concentration as the output. The ANN is able to correct mid-infrared (MIR) spectrometer measurements to within a margin of error of +/- 1% when compared to results from a relative humidity sensor (Bosch BME280). The key correction of the measurements is based on quantifying the amount of beam steering that occurs within the system. The system is a scanning grating-based MIR (2-10 micron) supercontinuum light source gas spectrometer. This spectrometer incorporates a 10 m pathlength and in its standard operation, can result in as much as 40% uncertainty due to ambient temperature fluctuations.
Journal Article Source
Sensors (Published 3 October 2023)
Abbreviations
ANN: artificial neural network
MIR: mid-infrared
Opinion
The researchers demonstrate a strong correlation between ambient temperature changes and a shift in the angle of incidence of the first order diffracted light that is focused on a single pixel detector in the spectrometer system. The °/K becomes a key parameter in the ANN. They also parametrize the least-mean-square noise from the data and the baseline noise that results from fitting of the baseline.
Small changes in ambient temperature (1-2 °C) can make measurements in a typical spectrometer over a long period of time difficult. These small temperature changes result in micro adjustments in the optical components of the system. This results in a lot of manual adjustment of the baseline measurement in order to try and account for the beam steering the changes in ambient temperature causes. Depending on the amount of data you collect, this can take hours or days to arrive at a usable series of final measurements.
By gaining an understanding of the angular displacement, the researchers were able to translate angular deviation into a change in the spectral wavenumber (cm-1). A grating angle deviation of approximately 0.015° resulted in a spectral shift of about 0.76 cm-1.
By implementing the ANN, the researchers were able to match data from a relative humidity sensor quite well, within 1% (see below). This is really the key in using neural networks for improving sensor results. It requires quantitative knowledge of your noise source. This can be extremely difficult to quantify in real-world applications. Most times you have no clue where the noise is coming from. So if someone on your team has familiarity with implementing ANNs, and you know and can quantify your noise sources, this approach could prove valuable. It would be quite interesting to implement this on a commercial level. Shoot us an email if you see this implemented elsewhere in laser absorption spectroscopy!
