Times are displayed in (UTC-05:00) Central Time (US & Canada) Change

Testing SandAI’s Performance on Quartz Grains from Glacial and Volcanic Provenance

Session: Advances and Applications in Geochronology for Interpreting Stratigraphic and Basin Records (Posters)

Presenting Author:

Nicholas Wong

Authors:

Wong, Nicholas¹, Castle, Victoria², Grandfield, Taylor³, Frouin, Marine⁴

(1) Dougherty Valley High School, San Ramon, CA, USA; Department of Geosciences, Stony Brook University, Stony Brook, NY, USA, (2) Department of Geosciences, Stony Brook University, Stony Brook, NY, USA, (3) Department of Geosciences, Stony Brook University, Stony Brook, NY, USA, (4) Department of Geosciences, Stony Brook University, Stony Brook, NY, USA,

Abstract:

SandAI is a deep neural network developed by M. Hasson in 2024 to classify the depositional history of quartz grains based on scanning electron microscope (SEM) images. It analyzes surface microtextures, which form through processes such as mechanical abrasion and chemical weathering, to categorize grains as fluvial, aeolian, beach, or glacial. However, previous studies have shown that SandAI struggles with grains affected by overprinting or from depositional environments outside of these four categories.

Based on these findings, we hypothesized that SandAI may not be able to reliably distinguish between glacial and volcanic grains. These two environments are often difficult to differentiate even through visual inspection. To test this, we selected two glacial samples from Long Island (NY) and one volcanic sample from Black Mountain (CO) and analyzed them using SandAI. Our goal was to evaluate the model’s performance on depositional types it was not explicitly trained on and determine whether it can differentiate between glacial and volcanic grains. If it cannot, this would indicate the need to retrain or expand the model to improve its classification capabilities.

For each sample, about 35 quartz grains were isolated through wet sieving and chemical treatment. The grains were gold-coated and images taken in a SEM, then processed through SandAI. Grains from Hither Hills were predominantly classified as beach-derived (87%), as were grains from Montauk (80%). In contrast, grains from the Black Mountain sample were classified as 50% beach, 25% fluvial, and 25% glacial.

These results suggest that SandAI struggles to accurately classify samples with complex depositional histories. The model appears biased toward identifying high-energy depositional environments, potentially overlooking more recent, lower-energy depositional processes. Furthermore, SandAI returned ambiguous classifications for grains outside the four depositional modes it was trained on. This underscores the need for a broader training dataset, particularly one that accounts for overprinting and includes volcanic materials as a distinct class.

Future research will examine how depositional energy influences model performance, investigate the role of environmental overprinting on grain surface features, and work toward integrating tephra as a new classification category in SandAI. Overall, this study highlights the importance of critically assessing AI-driven tools when applied to geologically complex materials.