How to feed emission spectra into machine learning models

Abstract

Machine Learning models are used on a variety of input data like image, text, tabular and video data. When we are trying to apply machine learning models in the natural sciences we, however, come across specialized data types that differ in a variety of ways and therefore we have to take extra care. In our lab we use a technique called laser-induced breakdown spectroscopy (LIBS) where we are creating a plasma by shooting a powerful laser onto rock samples and then measure the light from the plasma with a spectrometer. The output of this experiment is an emission spectrum which serves as the input to a machine learning model allowing us to quantify the elements and their abundances in the sample (see Fig. 1). In this work we have a look at different normalization, scaling and standardization procedures; which data augmentation techniques are feasible and what differentiates the emission spectrum from seemingly similar data types like the time series or a plain list of numbers.