Earth atmosphere emulators could enable faster and more accurate weather predictions

Earth's atmosphere is a complex system that can be modeled using billions of 1-kilometer-square chunks, but simulating a decade into the future takes months on supercomputers. Researchers at Lawrence Livermore National Laboratory (LLNL) are developing emulators to speed up weather predictions and provide more accurate forecasts months to years in advance.

These emulators use machine learning, like deep neural networks, to mimic the physics of the atmosphere based on data from observations or high-fidelity models. By training emulators like STRATA on the outputs of a model like E3SM, scientists can predict atmospheric conditions without solving the physics equations. This speed is essential for capturing the chaotic nature of weather, which requires many simulations to build trust in the predictive system.

Emulators also offer the ability to test the impact of initial conditions, as even small changes in starting points can lead to vastly different predictions. However, obtaining accurate initial conditions is challenging, especially for deep ocean temperatures and hard-to-reach areas. The Genesis Mission aims to address this issue by using generative models, like MAGGIE, to create consistent and accurate initial conditions for predictive simulations.

Additionally, LLNL is developing a small-scale emulator called STITCH to improve the representation of cloud microphysics and small-scale turbulence, which currently generate significant uncertainty in simulations.