101-13 OneStrata: A foundation model for deep-time biodiversity and mass extinction reassessment

Abstract:

The canonical "Big Five" mass extinctions have long been regarded as the principal drivers of Phanerozoic marine biodiversity history. This paradigm, however, was built upon taxonomic range compilations and stage-level time bins that cannot resolve ecological dynamics operating on timescales of thousands to tens of thousands 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. Unlike conventional correlation methods such as CONOP, which formulate stratigraphic correlation as a combinatorial optimization problem whose computational cost becomes prohibitive when datasets exceed approximately 10,000 fossil species, even on modern supercomputers, OneStrata reformulates the problem as a representation-learning task. The model integrates within-section stratigraphic superposition, species co-occurrence patterns, temporal windows, and radiometric age constraints into a unified probabilistic framework that scales to global databases.

Applied to a global dataset of 31,044 stratigraphic sections, 183,514 species, and 2,845,815 fossil occurrences, supplemented by 449 high-precision ID‑TIMS dates, OneStrata reconstructs marine biodiversity at an imputed average temporal resolution of thousands of years across ~600 million years, from the Ediacaran to the early Neogene. The resulting biodiversity trajectory challenges the classical "Big Five" paradigm: only the end‑Ordovician and end‑Permian events emerge as profound global biodiversity collapses, whereas the Late Devonian, end‑Triassic, and end‑Cretaceous extinctions appear comparatively weak or diffuse.

Beyond revising the mass‑extinction narrative, this study demonstrates how AI foundation models can transform fragmented geological observations into a continuously improvable probabilistic representation of Earth history. OneStrata establishes a scalable temporal infrastructure that can integrate paleontological, stratigraphic, geochronological, and environmental data within a unified deep‑time reference system. The framework is designed as a general infrastructure for deep‑time Earth‑system analysis rather than a biodiversity‑specific application, opening a pathway toward a predictive digital Earth for investigating the long‑term co‑evolution of life and the planetary environment.

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.