A new weather model predicts the future of the atmosphere using raw sensor data without ever dividing the Earth into a grid.
Weather forecasting has relied on chopping the world into a massive 3D grid of boxes for nearly a century to solve complex fluid equations. Earth-o1 ignores this entire mathematical structure and learns the continuous physics of the atmosphere directly from scattered sensor readings. This model matches the accuracy of current top-tier forecasting systems while skipping the most computationally expensive steps. Breaking free from the grid allows for much faster and more flexible predictions of extreme weather events. It represents a total pivot in how we simulate planetary systems using raw information rather than rigid geometry. The result is a system that can run on much simpler hardware while providing the same level of safety and foresight.
Earth-o1: A Grid-free Observation-native Atmospheric World Model
arXiv · 2605.06337
Despite the unprecedented volume of multimodal data provided by modern Earth observation systems, our ability to model atmospheric dynamics remains constrained. Traditional modeling frameworks force heterogeneous measurements into predefined spatial grids, inherently limiting the full exploitation of raw sensor data and creating severe computational bottlenecks. Here we present Earth-o1, an observation-native atmospheric world model that overcomes these structural limitations. Rather than relyin