Simulating low-likelihood high-impact extreme weather events in a
warming world is a significant and challenging task for current
ensemble forecasting systems. While these systems presently use
up to 100 members, larger ensembles could enrich the sampling of
internal variability. Due to computational constraints, it is
infeasible to generate huge ensembles (comprised of 1,000-10,000
members) with traditional, physics-based numerical models. In
this paper, we replace traditional numerical simulations with
machine learning (ML) to generate hindcasts of huge ensembles. We
construct an ensemble weather forecasting system based on
Spherical Fourier Neural Operators (SFNO). The
ensemble represents model uncertainty through perturbed-parameter
techniques, and it represents initial condition uncertainty
through bred vectors, which sample the fastest growing modes of
the forecast. We show that the ML extreme weather forecasts are
reliable and discriminating.

We generate a huge ensemble (HENS),
with 7,424 members initialized each day of summer 2023, the
hottest summer in at least the last 2000 years. For extreme
climate statistics, HENS samples events 4σ away from the ensemble
mean. We use HENS to characterize the spread in atmospheric river
intensity and impacts in the Southern Hemisphere. We also apply HENS
to ascertain whether tropical cyclones obey conventional or anomalous
diffusion subject to large-scale eddies in the general circulation.