Improving the performance of artificial neural networks trained on synthetic data in gas spectroscopy – a study on two sensing approaches

Abstract Artificial neural networks (ANNs) are used in quantitative infrared gas spectroscopy to predict concentrations on multi-component absorption spectra. Training of ANNs requires vast amounts of labelled training data which may be elaborate and time consuming to obtain. Additional data can be gained by the utilization of synthetically generated spectra, but at the cost of systematic deviations to measured data. Here, we present two approaches to train ANNs with a combination of comparatively small, measured data sets and synthetically generated data. For the first approach a neural network is trained hybridly with synthetically generated infrared absorption spectra of mixtures of N2O and CO and measured zero-gas spectra, taken with a mid-infrared dual comb spectrometer. This improves the mean absolute error (MAE) of the network predictions from 0.46 to 0.01 ppmV and 0.24 to 0.01 ppmV for the concentration predictions of N2O and CO respectively for zero-gas measurements which was previously observed for training with purely synthetic data. At the same time a similar performance on spectra from gas mixtures of 0–100 ppmV N2O and 0 to 60 ppmV CO was achieved. For the second approach an ANN pre-trained on synthetic infrared spectra of mixtures of acetone and ethanol is retrained on a small dataset consisting of 26 spectra taken with a mid-infrared photoacoustic spectrometer. In this case the MAE for the concentration predictions of ethanol and acetone are improved by 45 % and 20 % in comparison to purely synthetic training. This shows the capability of using synthetically generated data to train ANNs in combination with small amounts of measured data to further improve neural networks for gas sensing and the transferability between different sensing approaches.