Data Set05 for "Supervised machine learning approach to iron meteorite classification using Random Forest: A predictive model to classify ungrouped irons"

Data Set05 for "Supervised machine learning approach to iron meteorite classification using Random Forest: A predictive model to classify ungrouped irons". A. F. Rogers[1*], L. Doucet[2], L. V. Forman[1, 3], K. Rankenburg[4], G. K. Benedix[1, 3, 5] [1]Space Science and Technology Centre, School of Earth and Planetary Science, Curtin University, Perth, WA, Australia[2]Earth Dynamics Research Group, School of Earth and Planetary Science, Curtin University, Perth, WA, Australia[3]Department of Minerals and Meteorites, Western Australian Museum, Perth, WA, Australia[4]John de Laeter Centre, School of Earth and Planetary Science, Curtin University, Perth, WA, Australia[5]Planetary Science Institute, 1700 East Fort Lowell, Suite 106, Tucson AZ 85719-2395, USA.*Corresponding author: Ashley Faith Rogers ([e-mail], [e-mail]) Abstract Iron meteorites provide insight into the formation processes of planetary cores from the early Solar System. Through further classifying individual irons into specific groups, which are assumed to represent a common parent body, greater information can be yielded into the formation processes of those bodies. The current classification system of iron meteorites includes 13 groups distinguished by trace element compositions and mineralogical properties (i.e., bandwidth size of the α-Fe,Ni-alloy, silicate inclusions, etc). Group sizes vary with only 6 members identified as IIG irons and 391 members in the largest group, the IAB-complex. One-hundred and fifty-nine irons (or > 11% of all iron meteorites) that cannot be grouped with the current classification scheme are labelled ‘ungrouped’. This research explores whether supervised machine learning algorithms offer insights into classification of the ungrouped irons, despite the heavily imbalanced group data used for algorithm training. A Random Forest classifier is presented in this work that achieved weighted F1-scores of ~91 – 97% of testing datasets. This outcome indicated that machine learning is possible for the iron meteorite dataset. The accuracy of the model ranges from ~90 – 98%. As a result, at least 59 ungrouped irons could be reclassified to preexisting iron groups, 16 may belong to small grouplets while 12 should remain ungrouped. This work further discusses the implications of these potential reclassifications and how machine learning may be applicable to other meteorite studies.