Junior research group: Labs & Techniques

Machine learning (ML) is a subfield of artificial intelligence and relies on learning algorithms from ideally labelled and large datasets. ML is in particular valuable for applications handling complex problems and datasets where traditional approaches perform weakly. Furthermore, ML techniques can be used to recognize structures and patterns in large datasets, which are difficult to assess for humans.

Besides several advantages of using spectroscopy for in-situ planetary exploration, the data analysis of techniques like LIBS is challenged with inherent effects known as matrix effects. These effects can impede unambiguous classification and quantitative predictions of elemental abundances based on LIBS data. We investigate how ML models can be best trained and combined to reduce uncertainties in the prediction of compositions and in the classification of geological targets such as rocks and soils. Furthermore, following recent trends of scientific machine learning (SciML), we evaluate how physical knowledge of the measurement method or of experimental factors can be incorporated into the training.

Another aspect is that techniques like LIBS and Raman spectroscopy are more and more often selected for the payload of extraterrestrial missions. Consequently, datasets are continuously growing which allows us to apply statistical methods in order to support scientific interpretations of the data.

Reference:

K. Rammelkamp, O. Gasnault, O. Forni, C.C. Bedford, E. Dehouck, A. Cousin, J. Lasue, G. David, T. S. J. Gabriel, S. Maurice, R. C. Wiens, "Clustering Supported Classification of ChemCam Data From Gale Crater, Mars.", Earth and Space Science 8, e2021EA001903 (2021).

Data fusion of spectroscopic data

Spectroscopic data from techniques like LIBS, Raman and NIR spectroscopy give complementary information. While LIBS can reveal elemental compositions, Raman and NIR spectroscopy provide molecular information. If measured on the same sample, the combination of different spectroscopic techniques can therefore give a deeper understanding of the sample.

Often, spectroscopic instruments on robotic exploration missions also have a camera to capture the geological context. Such images can contain relevant information about the target such as the texture and grain size which can serve as additional information for classification and regression models. We investigate how different spectroscopic data can be best combined with each other but also with visual information and how ML algorithms can be trained on this multi-modal data to obtain the largest scientific return.

Reference: