MACHINE LEARNING APPLIED TO A PETROGRAPHIC DATABASE: INTRODUCING THE GLOBAL PREDICTION OF SAND MINERALOGY (GLOPRSM) MODEL

Sandstone petrography has long been used as a tool to infer sedimentary provenance. Petrographers of the 20^th century made significant advancements in relating the relative abundance of framework grains in sedimentary deposits to tectonic setting. Mineral proportions may provide insight to the boundary conditions (e.g. source terrane lithology, climate) of the systems in which sediments form. This research seeks to answer the following question: can the final modal composition of sand be predicted if boundary conditions are known? We investigate this question by analyzing a global compilation of published compositional data from Pleistocene-to-modern sand samples whose provenance and boundary conditions are known. Ultimately, this research aims to better resolve how Earth-surface processes are manifested in the global sedimentary archive.

To date, point count data for 3,512 sand samples have been compiled from 50 published sources. Environmental data were extracted from a subset 3,208 fluvial and marine sample catchments, including precipitation, temperature, relief, slope, basin area and source lithology. These data were used to train two Random Forest models to predict the log-ratios of samples’ quartz-feldspar-lithic (QFL) proportions. Preliminary results show 68% and 78% of the variability of the two log-ratios can be explained by the random forests. Using the BasinATLAS dataset, a Global Prediction of Sand Mineralogy (GloPrSM) was generated for level 8 watersheds (mean area ~8,000 km²). In general, GloPrSM predicts quartz enrichment in tropical latitudes (30°N to 30°S), feldspar enrichment near plutonic and metamorphic crystalline terranes, and lithic enrichment near active margins and flood basalts. Feature importance algorithms revealed that slope, temperature, metamorphic source abundance, and felsic to intermediate plutonic source abundance are the most important predictors of the log-ratio models. Future research will increase the granularity of GloPrSM by incorporating additional grain types, including mono- and polycrystalline quartz, chert, alkali feldspar, plagioclase, and volcanic, sedimentary and metamorphic lithics.