Sensitive and accurate determination of nitrogen in simulated Martian soil and environment with LIBS spectrum fusion and regression based on neural network

Abstract

Nitrogen is a key element necessary for the emergence and development of life. It is one of the elements targeted by the landed missions on Mars in accordance with their scientific goal of investigation of habitability and search for traces of life. A gas chromatography mass spectrometer (GCMS) instrument on board the Mars Science Laboratory Curiosity rover has revealed the existence of oxidized nitrogen-bearing compounds on Martian surface with an equivalent nitrogen concentration up to 0.01 wt%. Although the detection with laser-induced breakdown spectroscopy (LIBS) also on board the Curiosity rover is desirable, the current performance of LIBS for nitrogen analysis does not show the capacity in terms of limit of detection (LOD) and accuracy. Research on a suitable method should therefore be first engaged in laboratory in order to guide further improvements of LIBS instrument on board Mars rover, as well as the data treatment method. Beyond the sensitivity issue, matrix effect also affects LIBS determination of nitrogen, especially due to its various chemical speciation in geological materials. Method research should answer thus double requirements of improving the sensitivity and reinforcing the robustness with respect to different nitrogen-bearing compounds. An experimental configuration of double detections with a narrow bandwidth Czerny-Tuner (CT) spectrometer and a broad bandwidth Echelle spectrometer, was implemented in this work, in such way that the first ensured a sensitive detection of emission lines from nitrogen and the second complemented with those from major elements in the sample. The fusion of the simultaneously acquired spectra took into account the emission characteristics of the both two types of elements, necessary for an effective and robust multivariate regression based on a neural network. In addition, for a better treatment of different chemical speciation of nitrogen in samples, generalized spectrum was used for training of regression models, after an unsupervised clustering having assigned a type label to each training spectrum. The trained model was tested by collections of independent test samples, resulting in a limit of detection (LOD) of 0.18 wt%, and a root mean square error of prediction (RMSEP) of 0.041 wt%, representing a step forward to nitrogen detection using LIBS on Mars.