For generations, the evolutionary story of grasses — plants that now sustain much of human civilization — lay locked inside fossils too similar for human eyes to distinguish. Scientists at the University of Illinois Urbana-Champaign have now combined super-resolution microscopy with machine learning to read that story for the first time, tracing 25,000 years of grassland change on Mount Kenya and revealing how ice ages and rising carbon dioxide shaped the plants that shaped us.
Machine learning unlocks 25,000 years of grass evolution hidden in pollen
Millions of fossils, but no way to tell them apart
Mark
So the core problem was that grass pollen all looks the same. How much of a limitation was that really?
Mimi
Imagine having millions of documents in a library but no way to sort them. You know they contain information, but you can't access it. That's what paleobotanists faced. A single cubic centimeter of sediment could hold millions of pollen grains, but analyzing them was essentially impossible.
Luke
Wait—they had the pollen all along. They just couldn't read it. So this isn't discovering new fossils; it's developing a tool to interpret fossils that already existed.
Mimi
Exactly. The data was always there. The bottleneck was human ability to distinguish one grain from another.
Mark
And the machine learning solved that by finding patterns humans couldn't see?
Mimi
Not patterns humans couldn't see—patterns that were too subtle and numerous for humans to classify consistently. Super-resolution microscopy revealed surface textures and wall thickness variations. The neural network learned to recognize those variations and apply that learning to new samples.
Luke
But the model was trained on identifiable species. How do we know it's actually learning something real about the grasses, and not just fitting noise?
Mimi
They tested it on samples with known compositions. The diversity estimates matched the actual diversity. That's the validation.
Mark
What surprised you most about the Mount Kenya data?
Mimi
That grass diversity collapsed so dramatically during the Last Glacial Maximum. And then the shift in C4 to C3 ratios—C4 grasses dominated the coldest period, then gradually lost ground as it warmed. That's a clear climate signal written in pollen.
Luke
Though they note the C4-C3 shift doesn't have an obvious association with CO2 or temperature. So we're still not entirely sure why that happened.
Mimi
True. The mechanism isn't clear. But the pattern is now visible, and that's the first step.
Mark
What comes next?
Mimi
Refining the method, expanding it to other plant types. Essentially, unlocking 25,000 years of climate and ecosystem history that was locked in sediment cores all over the world.
El Pulso
- Millions of grass pollen fossils sat unreadable in sediment cores for decades because the grains are nearly indistinguishable under conventional microscopes — an entire field of evolutionary inquiry was effectively frozen.
- Super-resolution laser microscopy exposed subtle surface textures and cell wall differences invisible to the human eye, giving researchers their first real foothold into the problem.
- A convolutional neural network trained on those microscopic patterns can now estimate species diversity across mixed samples without identifying every individual grain — turning a human bottleneck into a scalable process.
- Applied to a 25,000-year sediment core from Mount Kenya, the method revealed that grass diversity collapsed during the Last Glacial Maximum and that C4 grasses — adapted to heat and drought — dominated until warming temperatures allowed C3 species to reclaim ground.
- Researchers are already extending the approach to other plant pollen and spores, suggesting that vast archives of fossil material long considered unreadable may soon yield their histories.
For generations, the evolutionary story of grasses — plants that now sustain much of human civilization — lay locked inside fossils too similar for human eyes to distinguish. Scientists at the University of Illinois Urbana-Champaign have now combined super-resolution microscopy with machine learning to read that story for the first time, tracing 25,000 years of grassland change on Mount Kenya and revealing how ice ages and rising carbon dioxide shaped the plants that shaped us.
For decades, paleobotanists faced a stubborn impasse: grass pollen grains look nearly identical under a microscope. A single cubic centimeter of lake sediment might hold millions of these fossils, yet scientists had no reliable way to tell them apart. Surangi Punyasena of the University of Illinois Urbana-Champaign and her former doctoral student Marc-Élie Adaimé recognized that this limitation was sealing off the evolutionary history of grasses — plants that feed much of the world and were among the first to be domesticated some 12,000 years ago.
The breakthrough came through super-resolution microscopy, which uses lasers and algorithmic reconstruction to reveal fine surface details invisible to conventional lenses. Examining grass pollen under this magnification, Adaimé noticed subtle variations in surface patterning and cell wall thickness that previous researchers had overlooked. Human eyes could not classify these differences consistently, but a machine could learn to.
Adaimé trained a convolutional neural network — a computational structure loosely modeled on how the brain processes visual information — to recognize these hidden patterns. He also developed a statistical method capable of estimating species diversity in mixed samples without identifying each grain individually. The model could even distinguish between C3 and C4 grasses, two fundamentally different photosynthetic types with contrasting advantages across climates.
Testing the method on a sediment core from a Mount Kenya lake, the researchers read 25,000 years of grassland history in microscopic detail. Grass diversity collapsed during the Last Glacial Maximum, when carbon dioxide levels were at their lowest and cold was most severe. As temperatures and CO2 rose in the millennia that followed, diversity recovered. C4 grasses — better suited to hot, dry conditions — were proportionally higher during the ice age's final phase before C3 species expanded with the warming climate.
The deeper significance lies in what the method unlocks going forward. Paleobotanists have always possessed enormous quantities of pollen fossils; they simply lacked the tools to extract meaning from them at scale. A trained neural network can now process thousands of grains in the time a researcher might examine dozens by hand. With plans to extend the approach to other plant pollen and spores, the deep evolutionary record of grasslands — an ecosystem present on Earth only since the Eocene, roughly 40 million years ago — can at last be read with both precision and speed.
For decades, paleobotanists have faced a frustrating problem: grass pollen grains look nearly identical under a microscope. A cubic centimeter of lake sediment might contain millions of these fossils, but scientists had no reliable way to tell them apart. The shapes, spikes, and grooves that make pollen from other flowering plants distinguishable simply don't exist on grass pollen. Surangi Punyasena, a plant biology professor at the University of Illinois Urbana-Champaign, and her former doctoral student Marc-Élie Adaimé recognized that this limitation was blocking an entire field of inquiry—the evolutionary history of grasses themselves, plants that now feed much of the world and were among the first to be domesticated roughly 12,000 years ago.
The researchers had already made progress using super-resolution microscopy, a technique that uses lasers to illuminate single points and algorithmically reconstructs scattered light to reveal fine surface details. When Adaimé examined images of grass pollen under this magnification, he noticed something previous researchers had missed: subtle variations in the patterning of the grain's surface and differences in cell wall thickness. These distinctions were too fine for human eyes to classify consistently, but they were there.
Adaimé trained a machine-learning model on images from several identifiable grass species, teaching it to recognize these hidden patterns. He then developed a statistical method that could estimate species diversity in samples containing pollen from multiple species—without needing to identify each individual grain. The model used convolutional neural networks, computational structures loosely inspired by how the human brain processes visual information. When tested on samples with known compositions, the method's estimates closely matched reality. The approach could also distinguish between C3 and C4 grasses, two fundamentally different photosynthetic types. C4 grasses concentrate carbon dioxide more efficiently, giving them advantages in hot, dry conditions, while C3 grasses dominate cooler environments.
With their method validated, Adaimé and Punyasena analyzed a sediment core from a lake on Mount Kenya representing 25,000 years of pollen accumulation. The results told a story written in microscopic detail. During the Last Glacial Maximum—the coldest, harshest stretch of the ice age between roughly 21,000 and 18,000 years ago—grass diversity plummeted. The atmosphere held extremely low levels of carbon dioxide. As temperatures rose and CO2 increased in the millennia that followed, diversity recovered. The proportion of C4 grasses showed a different pattern: higher during the final phase of the ice age, then gradually declining as the climate warmed and C3 species expanded their range.
What makes this work significant is not just the data it recovered, but the door it opens. Paleobotanists have always had access to vast quantities of pollen fossils; they simply lacked the tools to extract meaning from them. Human analysis was the bottleneck. Now, a computer trained on subtle visual patterns can process thousands of grains in the time it would take a researcher to examine dozens by hand. The method is already being refined, and researchers are planning to extend it beyond grasses to other plant pollen and spores. For the first time, the deep evolutionary history of grasslands—a relatively recent ecosystem in Earth's history, present only since the Eocene roughly 40 million years ago—can be read from the fossil record with precision and speed.
Citas Notables
Within a small cubic centimeter of sediment, you could have thousands, potentially millions of pollen fossils. But the level at which we were able to analyze it before machine learning was limited by human ability.— Surangi Punyasena, University of Illinois Urbana-Champaign
It is satisfying to see that there is so much more information to be unlocked. Now we have a way to begin unraveling the history of grasslands, and perhaps even the deep evolutionary history of grasses, from clues hidden in subtle differences among their pollen grains.— Marc-Élie Adaimé, Smithsonian Institution
Quieres la nota completa?
Lee el original en Phys.org ↗