Ammonite fossils preserved in limestone. These fossil deposits help scientists understand biodiversity throughout Earth’s history. Credit: Herve Conge/Science Photo Library Large language models (LLMs), such as OpenAI’s ChatGPT, are trained to recognize patterns in written text, helping them choose a statistically likely word to complete a sentence. Paleontologists have now trained a similar model with fossil records instead of words, and then used it to reconstruct a timeline of changes in Earth’s biodiversity over hundreds of millions of years in unprecedented detail. This big data method of interpreting the rock record (the field known as stratigraphy) could stimulate a rethinking of major upheavals in the history of life, the model’s developers say. For example, the model suggests that three of the “Big Five” mass extinction events were gradual declines in biodiversity, rather than the sudden cataclysms described by other reconstructions. A supercomputer scours the fossil record for hidden Earth extinctions. The model is “a real revolution for studying stratigraphy and past changes in diversity,” says Shuzhong Shen, a paleontologist at Nanjing University in China, who presented the research at a meeting of the Geological Society of America in Denver, Colorado, on December 12. OctoberOther researchers say that although the reconstruction of the model, called OneStrata, appears to be the most detailed global record yet, scientists already have high-resolution reconstructions of changes in biodiversity in specific places, including many that shed light on mass extinctions. “We could do some additional things with it, but I don’t know if it would be devastating to the Earth,” says Matthew Clapham, a paleontologist at the University of California, Santa Cruz. Researchers can piece together the grand story of adaptation and extinction throughout Earth’s history by marking the depth at which species are found in layers of sedimentary rock, called strata; Depth corresponds to the moment a fossil was deposited in geological time. However, it is a difficult computational problem to combine sequences from many different sites involving tens of thousands of species into a single coherent timeline of biodiversity change. To simplify this task, stratigraphers divide fossil sequences for each species into time intervals of millions of years. This makes it easier to assemble many sequences into a unified whole. But it produces a blurry view, obscuring changes in biodiversity that occur over shorter time intervals. For the past two decades, Shen and his colleagues have worked to get a sharper picture by compiling a huge dataset of fossils and other geological data called OneStratigraphy. This includes depth data from millions of fossil specimens in rock sequences from tens of thousands of sites around the world. Large fossil model In 2020, Shen and his colleagues ran an algorithmic method on a supercomputer to reconstruct thousands of these sequences, allowing them to estimate when a species appeared and went extinct with a resolution of 26,000 years, much higher than the resolution of previous estimates. They could then reconstruct a high-resolution timeline for 11,000 marine species over about 300 million years1. However, including more species in shorter intervals requires an impractical amount of computational resources. Biodiversity theory supported by island bird data. To get around that computational limit, Shen’s colleagues at the Zhejiang Laboratory (an AI-focused research institution in Hangzhou, China) used data from OneStratigraphy to train a machine learning model on statistical patterns in the fossil record. Just as an LLM uses its training to predict the next word in a sentence, the OneStrata model uses those patterns to predict when a particular fossil is most likely to have been deposited.
