Date of Graduation

12-2021

Document Type

Thesis

Degree Name

Master of Science in Geology (MS)

Degree Level

Graduate

Department

Geosciences

Advisor/Mentor

Glenn R. Sharman

Committee Member

Eugene Szymanski

Second Committee Member

Matthew D. Covington

Third Committee Member

Xiao Huang

Keywords

Petrography, quartz, feldspar, lithic grains

Abstract

Petrography has long been used as a tool to decipher the sedimentary provenance of sand and sandstone from the relative proportions of framework grain types. Petrographers have also related the proportions of quartz (Q), feldspar (F), and lithic (L) grains to the processes that form and modify sediments within sediment routing systems. This past work has shown that factors including source lithology, climate, transport history, and tectonism work in concert to modify the framework mineralogy of sand. However, there is a lack of a quantitative understanding of the interactions and feedbacks between these factors and how they modify sand mineralogy. This research aims to establish a predictive framework that constrains the relationship between sand framework grain mineralogy and the factors that influence it, including bedrock lithology, topography, and climate. Specifically, this study asks, “to what degree can the final modal composition of sand be predicted if the boundary conditions that generate sediments are known?”.

This question is investigated by analyzing a globally extensive modal point count dataset of 3,522 Pleistocene to modern sand samples from 51 published sources. A petrographic data model was created to standardize 287 reported petrographic labels to a final list of 54 labels. An inline series of random forest (RF) machine learning algorithms were trained on a subset of 3,208 fluvial and marine samples whose boundary conditions are known with a high degree of confidence. Data for precipitation, temperature, elevation, slope, basin area, and seven generalized source lithologies were extracted from sample catchments and used to train 100 RF meta-estimators that predict the logarithms F:Q and L:Q ratios as well as eight Q-F-L subcompositions, resulting in R2 scores of 0.654 ± 0.031 (1-sigma) and 0.706 ± 0.023 (1-sigma) for ln(F/Q) and ln(L/Q) models, respectively. Mean Q-F-L prediction error within one standard deviation is 2.6% ± 15% for Q, -1.1 ± 9.4% for F, and -1.5% ± 15.4% for L.

The Global Prediction of Sand Mineralogy (GloPrSM) model was generated by applying the 100 RF meta-estimators to a global dataset of fluvial watersheds (mean area of ~1,500 km2). The resulting Q-F-L prediction includes an estimate of spatial uncertainty based upon variability in the 100 predictions. In general, the GloPrSM model predicts quartz enrichment in low latitudes (35°N to 35°S), feldspar enrichment near plutonic and metamorphic crystalline terranes in middle to high latitudes, and lithic enrichment near active margins and flood basalts. Low model confidence is exhibited in catchments draining large igneous provinces, in sedimentary terranes in middle to high latitudes, and in orogenic settings. Feature importance algorithms reveal that slope, temperature, metamorphic source abundance, and felsic to intermediate plutonic source abundance are the most important predictors of Q-F-L composition. In addition, partial dependence analysis suggests temperatures higher than 15 ºC and large drainage areas favor quartz enrichment, while steeply sloping environments favor lithic enrichment. The GloPrSM model represents the first, global-scale estimate of sand mineralogical proportions, and illustrates that the spatial distribution of Q-F-L at Earth’s surface can be predicted from the first-order factors that generate sediments.

Share

COinS