A new transformer-based AI model trained on global fossil data suggests that three of Earth’s canonical big five
mass extinctions unfolded as gradual declines over millions of years rather than sudden cataclysms, challenging long-held evolutionary timelines.
For decades, the story of life on Earth has hinged on five major extinction boundaries. These sudden, devastating chapters wiped out vast portions of global biodiversity, clearing the slate for new eras to take hold. But a machine-learning approach applied to the rock record is prompting a major rethink of those evolutionary upheavals.
Researchers unveiled the new model—named OneStrata—at a meeting of the Geological Society of America in Denver, Colorado. Developed to interpret the vast physical history preserved in sedimentary rock layers, the model applies large-language-model techniques to geological data rather than written text.
How OneStrata Trains on Millions of Fossil Records
Just as artificial intelligence models learn statistical patterns in language to predict upcoming words, OneStrata trains on the fossil record to calculate when specific specimens were likely deposited. The system relies on a massive underlying dataset known as OneStratigraphy, which catalogs the depth and location of millions of fossil specimens recovered from tens of thousands of sites worldwide.
Traditionally, stratigraphers tackle the immense computational burden of combining scattered rock sequences by grouping fossils into broad time intervals spanning millions of years. While this strategy makes large datasets manageable, it creates a blurred timeline that obscures ecological shifts happening over shorter durations.
Researchers have spent two decades trying to sharpen that view. In 2020, scientists used supercomputer algorithms to analyze 11,000 marine fossils across thousands of stratigraphic sections, pushing time resolution down to roughly 26,000 years. However, scaling that algorithmic method to include more species at even finer intervals demanded impractical computing power. To bypass that limit, artificial intelligence researchers at the Zhejiang Lab in Hangzhou, China, trained the OneStrata machine-learning model on the OneStratigraphy dataset.
Neural Network Shows Three Mass Extinctions Were Gradual
Equipped with this high-resolution neural network, researchers used it to predict the placement of nearly four million marine fossils. The resulting analysis yielded a time resolution of just over 1,000 years, spanning more than 110,000 species over the past 540 million years.
Vast new fossil record challenges story of Earth’s ‘big five’ mass extinctions膨大な化石記録が地球の五大大量絶滅の通説に疑問を投げかける…
The output challenges conventional views of ancient biodiversity drops. In lower-resolution timelines, major extinctions appear as sheer cliffs. In the OneStrata reconstruction, three of the canonical big five
events—specifically the Late Devonian, end-Triassic, and end-Cretaceous extinctions—appear as prolonged, gradual declines stretching across millions of years.
Here we present OneStrata, a transformer-based AI foundation model that learns a shared probabilistic representation of geological time directly from fragmented stratigraphic observations.
Authors, via Yahoo
Shuzhong Shen, a palaeontologist at Nanjing University in China, notes that the findings do not necessarily mean these mass extinctions never occurred. Instead, the data may reflect a continuous turnover where dying species were rapidly replaced by emerging ones.
Scientific Caution and the Future of Stratigraphy
Shuzhong Shen described the framework as a real revolution for studying stratigraphy and past diversity change during his presentation in Denver.
Other specialists urge caution before rewriting textbooks. Matthew Clapham, a palaeontologist at the University of California, Santa Cruz, observed that while the model provides an extraordinarily detailed global overview, researchers already possess high-resolution local records for specific extinction intervals. Erin Saupe, a palaeontologist at the University of Oxford, expressed wariness about interpreting temporal patterns too literally, noting that apparent shifts in biodiversity could stem from underlying sampling biases in the fossil record.
Far from meaning that stratigraphy will disappear, this will free geologists to sort out more challenging questions about the history of the Earth.
Douglas Erwin, Santa Fe Institute
Douglas Erwin of the Santa Fe Institute suggests that these computational tools will complement traditional methods, allowing scientists to focus on the most complex mysteries of our planet's geological past.
