Researchers from the University of Stanford have developed an artificial intelligence-based tool—dubbed SandAI—that can reveal the history of quartz sand grains dating back hundreds of millions of years. With SandAI, researchers can tell with high accuracy if wind, rivers, waves, or glacial movements shaped and deposited motes of sand. 

The tool gives researchers a unique window into the past for geological and archaeological studies, especially for eras and environments where few other clues, such as fossils, are preserved through time. SandAI's approach, called microtextural analysis, can also help with modern-day forensic investigations into illegal sand mining and related issues.

"Working on sedimentary deposits that haven't been disturbed or deformed feels about as close as you can get to being in a time machine—you're seeing exactly what was on the surface of Earth, even hundreds of millions of years ago. SandAI adds another layer of detail to the information we can pull from them," said Michael Hasson, a PhD candidate with Mathieu Lapôtre, an Earth and planetary sciences assistant professor at the Stanford Doerr School of Sustainability.

AI-powered tool

Historically, microtextural analysis has been done by hand and eye, using magnifying glasses and microscopes to attempt to draw inferences about sand grains' histories.

Modern science has validated the approach, showing that transport mechanisms do indeed impart telltale signatures. For example, grains that travelled farther often appear more rounded because their sharp corners have dulled; waves and wind also leave distinctive abrasion patterns.

However, traditional microtextural analysis is highly subjective, time-consuming, and scattershot across different studies. Thanks to the new tool, which leverages the power of machine learning to deeply scrutinize microscopic images of sand grains, microtextural analysis can now be far more quantitative, objective, and potentially useful across a wide range of applications. It also analyzes individual sand grains instead of lumping multiple grains into a single category, offering a more complete evaluation.

Worldwide, sand is the most used resource, after water, and is critical in the construction industry. Materials such as concrete, mortar, and some plasters require angular sand for proper adhesion and stability. However, gauging the origins of sand to ensure ethical and legal sourcing is challenging, so the researchers hope SandAI can bolster traceability. For example, SandAI could help forensics investigators crack down on illegal sand mining and dredging.

To build SandAI, the researchers employed a neural network that "learns" in a manner akin to the human brain, where correct answers strengthen connections between artificial neurons, or nodes, in the program, enabling the computer to learn from its mistakes.

With help from collaborators around the world, Hasson assembled hundreds of scanning electron microscope images of sand grains, representing material from the most common terrestrial environments: fluvial (rivers and streams), eolian (windblown sediments, such as sand dunes), glacial, and beach. 

SandAI analyzed this set of images to train itself to predict the sand grains' histories based on features that human researchers might never discern. The tool naturally made errors and would then iteratively improve. Once SandAI reached a robust 90% prediction accuracy, the researchers introduced new samples the model had not previously seen.

Novel applications

SandAI surmised that the ancient sand grains had been shaped and deposited as part of a windblown sand dune, in agreement with some manual microtextural studies. Moreover, because the tool analyzes individual sand grains rather than lumping multiple grains into a single category, other details emerged.

While the dominant signature indicated wind transport, a secondary signature that manual techniques would likely miss pointed to glacial sand. Together, those signals paint a portrait of dunes running somewhere near a glacier, as might well be expected during the Snowball Earth period.

The researchers have made SandAI available online for anyone to use. They plan to continue developing it based on user feedback and look forward to seeing the tool applied in a range of contexts.

Phys Org