Earth Atmosphere Emulators Enable Faster, More Accurate Weather Predictions

Lawrence Livermore National Laboratory researchers are developing machine-learning emulators designed to make long-range weather prediction faster, broader and more practical for national security planning. Their work centers on replacing, or augmenting, parts of the traditional physics-based Earth system modeling process with trained surrogate models that can approximate complex atmospheric outcomes without solving costly equations cell by cell. The primary effort is the STRATA emulator, which is trained on outputs from models such as the U.S. Department of Energy’s Exascale Earth System Model, or E3SM. Conventional Earth system simulations require high-performance supercomputers to run billions of grid-cell calculations, making decade-scale forecasts extremely time-intensive. STRATA instead uses deep neural networks to learn how the atmosphere evolves, allowing scientists to produce future states over periods ranging from days to years in a fraction of the required compute time. The laboratory says the speed advantage is not simply a convenience. Weather systems are chaotic, so credible forecasting requires running hundreds of simulations rather than relying on a single result. Emulators make those large ensembles feasible, giving researchers probability-based assessments instead of isolated predictions. That capability could support protection of critical infrastructure, supply-chain resilience, disaster preparation and military or naval planning. The work also addresses a key weakness in weather modeling: uncertainty in initial conditions. Because short-term forecasts depend heavily on the starting state of the atmosphere and oceans, poor or incomplete observations can distort results. A separate Department of Energy Genesis Mission project, MAGGIE, led by Los Alamos National Laboratory, is developing generative methods to create better initial conditions for E3SM-based simulations. Lawrence Livermore will help test those conditions using STRATA, a task that would have taken about three months with physics-based models but was completed over a single weekend using emulators. A second frontier is cloud physics. Emulators can at best match the skill of the model on which they are trained, so researchers are building more specialized surrogates to correct errors in small-scale processes. The STITCH project focuses on cloud microphysics and turbulence, phenomena that are difficult to measure and represent in global models. By training emulators on localized simulations of droplet formation, collisions, phase changes and turbulent motion, the team aims to create more faithful parameterizations that can be incorporated back into E3SM. Those process-level emulators are being developed using Lawrence Livermore’s Energy Research and Forecasting model with advanced super droplet cloud microphysics. Outputs will be validated against observational data from Department of Energy Atmospheric Radiation Measurement facilities, including deployment-site emulators that help connect model results with real-world measurements. Over time, improved cloud physics in E3SM can generate higher-quality training data for STRATA, closing the loop between microscale processes and global forecasting. The broader goal is to build emulators that both accelerate and improve Earth system predictions, linking meter-scale cloud behavior to global weather outcomes. Remaining challenges include generating sufficient training data and combining methods across physics, machine learning and high-performance computing. Still, researchers say the approach could materially strengthen forecasts of water resources, extreme precipitation, drought stress and other events with national security and economic significance.