PREDICTING SAND MINERALOGY WITHIN CORDILLERAN OROGENIC SYSTEMS: MACHINE LEARNING APPLIED TO A GLOBAL DATABASE OF MODERN-PLEISTOCENE SAND SAMPLES (Invited Presentation)
This presentation explores a new global prediction of sand mineralogy (GloPrSM) with an emphasis on the North and South American cordillera. GloPrSM is based on paired random forest models that predict the total abundance of quartz (Q), feldspar (F), and lithics (L), along with sub-grain types, for level 8 watersheds (~8,000 km2 mean area) of the BasinATLAS dataset. The GloPrSM model is calibrated using modal point count data from >3,200 modern-Pleistocene sand samples compiled from over 50 published sources, in conjunction with known values of precipitation, temperature, relief, slope, basin area, and source lithology from upstream watersheds. The GloPrSM model is in broad agreement with expected provenance patterns predicted by Dickinsonian Q-F-L ternary provenance fields. However, pronounced quartz enrichment within low latitudes (30°N to 30°S), likely the result of enhanced weathering of unstable siliciclastic grain types in the tropics, suggests significant climatic overprint on underlying tectonic regimes. Quartz enrichment at the expense of lithic and feldspar content is observed along many river profiles (e.g., the Amazon River), highlighting the complexity of sand’s compositional evolution within some sediment routing systems.