Neural Atmospheric Emulators
LLNL Trains STRATA Neural Emulator to Speed Global Weather Forecasts
Backed by the DOE Genesis Mission and NVIDIA, Livermore scientists Peter Caldwell and Yunyan Zhang replace expensive global physics equations with machine learning models trained on E3SM outputs.

Your listening Queue & catch up
Add a story or build a catch-up queue to start listening.
Computational climate scientists at Lawrence Livermore National Laboratory (LLNL) are replacing step-by-step atmospheric physics equations with deep neural network emulators that accelerate global weather simulations by orders of magnitude.
According to The Independent's report on the Nov. 4 LLNL announcement, a joint research team from LLNL, the U.S. Department of Energy (DOE), and NVIDIA is training machine learning emulators under the DOE's Genesis Mission—a federal initiative to build a national artificial-intelligence platform—to protect power grids, dams, and agricultural crops from extreme weather and assist military operations planning.
Because an air mass takes roughly a week to circulate around the Earth, accurate long-range forecasting requires full-globe simulations that are computationally intensive when solved via traditional fluid dynamics equations. LLNL scientist Peter Caldwell explained that the team's primary emulator, named STRATA, trains a deep neural network prediction engine directly on historical observations and high-fidelity outputs from the DOE's Exascale Energy System Model (E3SM). Given a set of initial atmospheric conditions, STRATA generates global precipitation and humidity forecasts orders of magnitude faster than physics-based runs while remaining nearly indistinguishable in output fidelity.
Complementing STRATA is a second Livermore effort led by LLNL scientist Yunyan Zhang: the Scalable Twin for Intelligent Turbulence and Cloud Heuristics (STITCH). Rather than emulating the entire globe at once, STITCH trains deep neural networks on fine-scale cloud microphysics and internal turbulence to replace traditional physical parameterizations inside larger climate models.
"The emulators for cloud microphysics and turbulence can be run hundreds or thousands of times faster than the high-fidelity physics models," Zhang said, noting that embedding STITCH process emulators inside host models improves physical representation accuracy at comparable or lower computational cost.
Sources
Follow the coverage
Published reporting connected by Machine Learning.
- Go.AI Raises $85M Series A as OpenAI Tests Wet-Lab Biology AI
- LLNL Trains STRATA Neural Emulator to Speed Global Weather ForecastsYou are reading this story
Comments (0)
No comments yet. Be the first to share a thought on this story.
Request a correctionContact the newsroom