REVEALING THE HIDDEN STRUCTURE OF THE LUNAR INTERIOR: INSIGHTS FROM MACHINE LEARNING (Invited Presentation)

The chemical, petrological, and physical structures of the lunar interior are such that individual modeling techniques serve as limited tools for deciphering these aspects of the lunar interior. A confounding factor to mantle modelling is that the generally accepted cumulate mantle overturn (CMO) induced late-stage mixing among two layers: the mantle and the immediately overlying (sub-crustal), denser Fe-Ti-rich cumulate layer. This mixing is hypothesized to be one of the two primary causes of mantle heterogeneity (the other being petrological stratification, a consequence of lunar magma ocean crystallization). The various methods that have been used to explore the structure of the lunar interior range from traditional approaches grounded in seismic data and geochemical/petrological analyses of returned lunar samples, to thermodynamic modeling and high P-T experiments simulating interior processes. Because no single method is capable of revealing a tightly constrained, unique solution for mantle geochemistry or mineralogy/petrology profiles, combined-method approaches are typically employed. We investigate the use of machine learning algorithms for gaining insight into the structure and composition of the lunar mantle. Using a newly complied database of Apollo basalt characteristics, nine distinct variables (seven basalt major element oxides, four mineral modes, textural categories, and age; n=207) were grouped on varying spatial scales and analyzed by k-means cluster analysis (KCA) and principal component analysis (PCA). The results show spatial-based variability in correlation strengths among the nine variables. For example, when the variables are treated as a single nearside group, a pattern emerges, suggesting moderate basalt coarsening with decreasing TiO₂/ilmenite content. This behavior is coupled with increasing Al₂O₃/plagioclase at the expense of FeO. The two behaviors suggest ilmenite-rich basalts may have experienced relatively increased undercooling while source magmas reflecting increased fractionation via increased FeO-enrichment do so with reduced plagioclase saturation. When the same variables are treated as two distinct groups (inside and outside the Procellarum KREEP Terrane (PKT), respectively), a more pronounced trend is expressed inside the PKT with respect to increasing Al₂O₃/plagioclase at the expense of FeO. These machine learning approaches, even if an assumption of Apollo sampling bias holds, demonstrate novel tools that can be used alongside more standard methods.