Super-resolution reconstruction of noisy gas-mixture absorption spectra using deep learning

•Gas-mixtures’ absorption spectra resolution improving using super-resolution (SR) reconstruction approach is proposed.•An artificial neural network based on multilayered perceptron (MLP ANN) demonstrated better spectral SR reconstruction compared to a convolutional neural network.•The two-stage and four-stage MLP ANNs were proposed for spectra SR reconstruction. In common, at low noise, the sequential models improve SR reconstruction quality compared to the single-stage MLP ANN. The laser sources used in absorption spectroscopy of gas media are a compromise between spectral turnability and line width. For example, optical parametric oscillators have a very wide tuning range, but also have a rather wide laser radiation linewidth. To mitigate the disadvantages of the latter, an approach to absorption spectroscopy gas-analysis spectral resolution improving using super-resolution (SR) reconstruction is proposed. It was implemented using several machine learning models based on different artificial neural network (ANN) architectures, including an original sequential ensemble ANN approach. The problem of random noise influence on SR reconstruction quality was resolved in two ways: (i) by learning convolutional neural networks with noisy spectra, (ii) by high-frequency noise preliminary decreasing, using Gaussian or Fast Fourier Transform filtering. The following ANN architecture models were designed and tested: convolutional neural network (CNN) and multilayer perceptron (MLP). The former performed at a lower accuracy compared to the latter. ANNs' sequential combination was implemented when each subsequent ANN used the results of previous ANN data processing. This architecture pursues a paradigm of ensemble algorithms. Sequential models consisting of two or five MLP ANNs were designed and tested. In general, at low noise, the sequential models provided better SR reconstruction quality compared to the single-stage MLP ANN. When the noise amplitude was 4% and more, the sequential models demonstrated 3–8% worse accuracy than the single-stage MLP ANN, even using filtering. Therefore, the sequential models are quite accurate and effective in combination with effective filtering in cases of moderate noise level. [Display omitted]