Global atmospheric data assimilation with multi-modal masked autoencoders

Global data assimilation combines observations and model outputs to determine the current state of the atmosphere. We present “EarthNet”, a multi-modal foundation model for data assimilation that predicts a global gap-filled atmospheric state solely from multi-modal satellite observations. EarthNet ingests a 12-hour sequence of observations and learns to fill missing data from other sensors. We show that EarthNet produces a global 0.16° atmospheric state of 3D atmospheric temperature and humidity 50x faster compared to operational systems (5 minutes on 1 GPU versus 4 hours on thousands of CPU nodes). The resulting 3D atmospheric dataset reproduces the climatology by evaluating a 1 hour forecast background state against observations. We also show that EarthNet's 3D humidity predictions outperform MERRA2 and ERA5 reanalyses by 10% to 60% between the middle troposphere and the lower stratosphere (5 to 20 km altitude) while underperforming near the surface and for 3D temperature. It is shown that our 3D temperature and humidity are statistically equivalent to the Microwave integrated Retrieval System (MiRS) observations at nearly every level of the atmosphere. Our results indicate significant promise in using EarthNet for high-frequency data assimilation and global weather forecasting.