Machine Learning Unlocks 25,000 Years of Grass Evolution Through Pollen Analysis

For decades, paleontologists studying ancient grasslands have faced a frustrating wall: grass pollen grains look nearly identical under a microscope. Unlike pollen from other flowering plants, which display distinctive shapes, spikes, grooves, or pore patterns that make identification straightforward, grass pollen presents a uniform face to researchers. This sameness has made it nearly impossible to distinguish one grass species from another in the fossil record, leaving a significant gap in our understanding of how grassland ecosystems have evolved over time.

A team led by Surangi Punyasena, a plant biology professor at the University of Illinois Urbana-Champaign, set out to solve this problem using tools that didn't exist when the challenge was first identified. Working with Marc-Élie Adaimé, a former doctoral student now at the Smithsonian's Office of Digital and Innovation, Punyasena developed a machine-learning approach paired with super-resolution microscopy to detect the subtle differences that human eyes and standard equipment cannot perceive. The method works by training artificial intelligence to recognize minute variations in pollen grain structure—variations too small or too fine to spot with conventional observation.

The payoff was substantial. Using their new technique, the researchers were able to analyze a single site's pollen record spanning 25,000 years, tracking how grass diversity shifted across a quarter of a million years of environmental change. This represents a genuine breakthrough in paleobotany, opening a window into ancient grassland composition that has remained largely closed. The work appeared in the Proceedings of the National Academy of Sciences, signaling its significance to the broader scientific community.

The stakes of this research extend far beyond academic curiosity about ancient plants. Grasses arrived relatively late in Earth's evolutionary history—they were present during the Eocene epoch, roughly 40 million years ago, when open grasslands began to emerge as a distinct ecosystem type. But their importance accelerated dramatically around 12,000 years ago, when humans began domesticating grass species. That domestication event transformed human civilization. Today, grasses provide some of the world's most critical food crops: wheat, rice, maize, barley, sorghum, and millet feed billions of people. Understanding how grass diversity has changed over millennia—how species composition shifted in response to climate, fire, and other environmental pressures—offers insight into the plants that now anchor global food security.

The machine-learning method Punyasena and Adaimé developed essentially teaches computers to see what humans cannot. By analyzing thousands of pollen grains and learning the subtle morphological signatures that distinguish one grass species from another, the algorithm can classify fossil pollen with a precision that was previously impossible. This capability doesn't just apply to a single location or time period. The technique can be applied to pollen cores and sediment samples from around the world, potentially reconstructing the natural history of grasslands across continents and millennia.

What emerges from this work is a new capacity to read the deep history written in pollen grains—to understand not just that grasses existed in the past, but which grasses, in what proportions, and how those communities changed. For researchers studying how ecosystems respond to environmental shifts, how human land use has reshaped plant communities, or how domestication altered the genetic and ecological landscape of our most important crops, this represents a genuine tool for discovery. The 25,000-year record from a single site is only the beginning. As the method spreads and researchers apply it to pollen archives around the world, the hidden history of grasslands—and the plants that feed us—will finally become legible.