Abstract
Extreme weather amplified by climate change is causing increasinglydevastating impacts across the globe. The current use of physics-basednumerical weather prediction (NWP) limits accuracy due to high computationalcost and strict time-to-solution limits. We report that a data-driven deeplearning Earth system emulator, FourCastNet, can predict global weather andgenerate medium-range forecasts five orders-of-magnitude faster than NWP whileapproaching state-of-the-art accuracy. FourCast-Net is optimized and scalesefficiently on three supercomputing systems: Selene, Perlmutter, and JUWELSBooster up to 3,808 NVIDIA A100 GPUs, attaining 140.8 petaFLOPS in mixedprecision (11.9%of peak at that scale). The time-to-solution for trainingFourCastNet measured on JUWELS Booster on 3,072GPUs is 67.4minutes, resultingin an 80,000times faster time-to-solution relative to state-of-the-art NWP, ininference. FourCastNet produces accurate instantaneous weather predictions fora week in advance, enables enormous ensembles that better capture weatherextremes, and supports higher global forecast resolutions.