Late Paleogene loess formed during a period of active regional tectonism and global climate cooling. Although Oligocene loess has been documented in central Asia and North and South America, it remains unclear whether the earliest appearances of loess were simultaneous across all regions before the Eocene−Oligocene cooling and whether they provided feedback on global climate during this time. This study examines the earliest appearance of late Paleogene loess at Flagstaff Rim in Wyoming, western United States, by integrating detailed sedimentologic characterization with both supervised and unsupervised machine learning models. Integrating grain-size patterns and field sedimentological features in the upper Eocene−Oligocene White River Formation led to the identification of four depositional processes, which were further supported by two unsupervised machine learning methods. The four patterns reflect primary loess, pedogenically modified loess, fluvial reworked loess and/or mixed dust-sand storm deposits, and fluvial sediments. Loess deposition occurred during active fluvial process as early as 35.8 Ma, 0.5 m.y. earlier than the apparent lithofacies change. By using the geological classifications as predefined labels, we trained four supervised machine learning models to identify loess from the grain-size data from other sites of the White River Formation (Group) in the western United States. Our multilayer perceptron (MLP) model achieved the best performance and predicted pedogenically modified loess and earlier starts of loess accumulation than previously thought at two of the sites. This study suggests that loess coverage in the western United States was significant before the drastic cooling at the Eocene−Oligocene transition, similar to loess in central Asia. Wind transport of loess and finer dust could have served as feedback on the latest Eocene global cooling. Additionally, this study demonstrates that using machine learning to analyze grain-size patterns not only helps interpret depositional processes but also identifies new features in the dataset that facilitate the formulation of new hypotheses for further investigation.
Guo et al. (Wed,) studied this question.