Automated curve fitting and unsupervised clustering of manganese oxide Raman responses

Natural manganese oxides characterization represents a challenge due to the broad variety of their structures and geochemical compositions along with a frequent poor crystallinity. This characterization requires the ability to conduct both a phase separation and a phase association operation. In this paper, the Raman spectra acquired on a selection of natural manganese oxide minerals are first processed with an automated curve-fitting model called MnOx. The adjustment of convolution envelope is realized, thanks to the Levenberg–Marquardt algorithm applied on a set of randomly generated seed pseudo-Voigt curves. The application conditions of the automated curve-fitting and, in particular, the number of seed curves are investigated with regards to the risk of overfitting. The MnOx model outputs are in a second step treated by data mining techniques and in particular by unsupervised clustering methods. This data processing shows promising results in terms of phase separation when the number of clusters is equivalent to the number of phases. By contrast, the decrease of number of clusters leads to a phase association which reflects spectral affinities between phases. This result shows the existence of 6 to 7 vibrational bands in Mn oxides Raman spectra, with contrasting behaviours between clustering. Thereby, vibrational bands located in the low wave number domain (<510 cm−1) are more mobile and therefore more sensitive to structural modifications, while bands with higher wave number are less affected by structural changes. Besides these results, Raman responses collected during this study provide new refinement regarding the spectral content of some Mn oxides in particular todorokite, nsutite, and chalcophanite.

Peer Review, Impact Factor

Models, mines, mineralogy