Marc-Élie Adaimé stared at a screen filled with tiny, golden pollen grains, each no wider than a tenth of a millimeter, and saw what no one else had: hidden patterns in plain sight. These specks — from grasses like Poa anceps and Andropogon amethystinus — looked nearly identical under ordinary microscopes, masking 25,000 years of evolutionary history. Now, using super-resolution microscopy and machine learning, Adaimé and his advisor, University of Illinois professor Surangi Punyasena, have cracked open a door long sealed to scientists.
Grasses feed the world. Wheat, rice, maize, barley, sorghum, millet — all come from this single plant family. Open grasslands only became widespread about 40 million years ago, but they now cover nearly 40% of Earth’s land. Yet for decades, researchers struggled to trace how these vital plants evolved, because their pollen is so hard to tell apart. Under standard light microscopes, grass pollen grains look like smooth, featureless spheres. Even electron microscopes could help only so much — they’re slow, costly, and impractical for analyzing thousands of fossils in sediment samples.
The breakthrough came when Adaimé used super-resolution microscopy, a technique that captures scattered light and reconstructs it into ultra-sharp images — almost as detailed as electron microscopy, but faster and easier. He spotted subtle differences in surface patterning, complexity, and cell wall thickness between C3 and C4 grasses — two major groups with different ways of photosynthesis. C4 grasses, which thrive in hot, dry climates, include corn and sorghum; C3 grasses include wheat and rice. Using over 1,000 high-res images, he trained a deep-learning model based on convolutional neural networks to recognize these tiny distinctions.
Then came the real test: Could the model estimate species diversity in mixed samples? Yes. When tested on known mixtures, its predictions matched actual compositions with remarkable accuracy. For the first time, scientists can now analyze fossil grass pollen not just by type, but by diversity and proportion — revealing how grass ecosystems changed through time. At a site in Central America, the team used the method to track shifts in grass types over 25,000 years, linking changes to ancient climate shifts.
This isn’t just about the past. Understanding how grasses responded to climate change before helps predict how crops might fare in a warming world. "That's essentially the definition of machine learning: for a computer model to be able to learn patterns and apply that learning to new data without being given explicit rules," Adaimé said. With this tool, paleobotanists are no longer blind to one of Earth’s most important plant families. The golden dust of ancient grasses now tells a richer, deeper story.