Huge Ensembles for Extreme Weather Prediction

Huge ensembles for extreme weather (HENS), a large collaboration between NERSC, LBNL EESA, NVIDIA, and Indiana University done as part of NERSC’s NESAP program has demonstrated an innovative new approach to ensemble weather forecasting that leverages deep learning models to produce much larger ensembles than what has been possible previously. Traditionally, numerical models require massive computational resources to simulate large ensembles of forecasts—key for accurately capturing the variability of rare but impactful weather phenomena. Current models typically use up to 100 members, but expanding ensemble sizes to thousands could significantly improve uncertainty quantification and understanding of the physical drivers of extreme weather. However, the computational burden makes this impractical with existing physics-based simulations. In a two-part study, the HENS project has demonstrated how machine learning, specifically deep learning emulators based on Spherical Fourier Neural Operators (SFNO), can generate ensembles of unprecedented size while overcoming these computational challenges.

In Part I of the study, the collaboration developed a machine learning-based forecasting system designed to replace conventional numerical simulations. This system captures both model uncertainty and initial condition variability using multi-model ensembling and bred vectors, respectively. By incorporating large-scale, distributed SFNOs with 1.1 billion learned parameters, the ensemble successfully passed critical qualitative and statistical tests, showing realistic weather states in individual forecasts as well as realistic ensemble spread as individual trajectories propagate forwards in time. The probabilistic forecasts generated by this system were found to be highly calibrated, performing comparably to the European Centre for Medium-Range Weather Forecasts IFS model in aggregate ensemble statistics as well as extreme event case studies like severe storms and heatwaves.

Part II expands the scope of the research, introducing a "huge ensemble" (HENS) of over 7,400 members, which was tested over the summer of 2023. This production run was generated on 256 Perlmutter GPUs, generating over 3PB of outputs to be analyzed. The HENS forecasts provide an unprecedented level of detail in sampling the tails of the forecast distribution, vastly improving the model's ability to predict extreme weather events and reducing the uncertainty of extremes by an order of magnitude. At each grid cell, HENS improves the skill of the most accurate ensemble member and enhances coverage of possible future trajectories; it also reduces the likelihood of outlier events where actual weather outcomes fall outside forecasted ranges. This breakthrough promises to enhance the precision of weather forecasts, offering a powerful tool to mitigate the societal impacts of extreme weather in an era of climate change.