TerraNova Trains on 1,024 Records in Native Geometries to Stop Averaging Earth Data Across Political Borders

arXiv 2607.29527·medium signalXBlueskyLinkedInCopy link

TerraNova argues the barrier to coupling physical-Earth and societal modeling is geometric, not statistical: the Earth is measured as continuous fields indifferent to borders while societies are reported per administrative unit, and prior work bridged the two by lossy averaging. It trains on 1,024 records in their native geometries — 512 gridded Earth-system fields plus 512 national indicators — with dedicated encoders for location, country, time, and task, cross-modal transformers fusing them, and a hypernetwork generating a per-query decoder with an evidential head returning predictive distributions. The frozen backbone reconstructs dense fields from sparse observations and adapts to unseen variables in minutes on consumer hardware.