Unsupervised machine learning with petrological database ApolloBasaltDB reveals complexity in lunar basalt major element oxide and mineral distribution patterns

Abstract

Diversity of lunar basalt characteristics is partly a consequence of lunar mantle heterogeneity. Although the cumulate mantle overturn hypothesis is the current standard model invoked to explain mantle asymmetries of unknown length scale in both compositional and geometrical space, successful petrological modeling of this mixing event requires a specific set of parameters not currently agreed upon. In contrast, surface basalt patterns may yield clues to both localized and nearside lunar interior structure.Using two multidimensional data analysis approaches – principal component analysis (PCA) and K-means cluster analysis (KCA) – we report the patterns produced from basalt characteristics over changing spatial scales, from intra-site to inter-site to nearside. The data are sourced from a newly developed, self-contained database of lunar basalt characteristics (ApolloBasaltDB), which includes major element oxides, mineral modes, ages, and textures for petrological and statistical modeling. Through the simultaneous considerations of multiple basalt characteristics contained in the database, we find that terrestrial-based basalt classifications cannot adequately describe the complex and overlapping distribution patterns of major element oxides and mineral modes that define multiple distinct basalt groupings over multidimensional space. These patterns provide opportunities for alternative lunar basalt classification schemes. Our analyses suggest that Al2O3 volumetric content is more diverse inside the Procellarum KREEP Terrane rift boundary versus content for older Apollo samples in close proximity to the eastern arm of the same rift boundary. Northernmost basalt samples show increased pyroxene diversity. Easternmost sites suggest anti-correlations in modal ilmenite and plagioclase, based on major element oxide PCA biplots, while nearside analyses of either major element oxides or mineral modes similarly suggest plagioclase (and Al2O3) diversity comes at the expense of ilmenite (and TiO2) diversity. There is evidence to suggest that approximate mineral content can be extracted from major element oxide data based on correlative patterns between major element oxide PCA biplots and mineral mode PCA biplots. These patterns have implications for remote sensing missions in that onboard data manipulation may provide lithologic basalt vectors of interest.