Sonny
A compute-light global weather backbone for medium-range forecasting. Sonny-ENS and KNOU-S2S grow from it, and Pyree borrows its backbone design.
5.5 daysto train to convergence on one NVIDIA A40 GPU, on a 1.5° global grid with 6-hourly steps.
Nine models for weather, climate and the hazards that follow from them. Several of the forecast models grow from one backbone, Sonny.
Forecast horizon on a log scale, from six hours to a century. Select a bar to jump to its entry.
A compute-light global weather backbone for medium-range forecasting. Sonny-ENS and KNOU-S2S grow from it, and Pyree borrows its backbone design.
5.5 daysto train to convergence on one NVIDIA A40 GPU, on a 1.5° global grid with 6-hourly steps.
A probabilistic ensemble of three Sonny-lineage models. A 30-day rollout takes 3.2 seconds per member on one A40.
20 of 20variable and lead-time scores with lower CRPS than the ECMWF IFS ensemble (2020, 18 members).
Subseasonal-to-seasonal prediction from a slow clock and two fast clocks that are fused into one ensemble.
48 of 48global RMSE and ACC comparisons beating the IFS extended-range ensemble mean, weeks 3 to 6 (2022 issue dates).
A small corrector that turns a frozen neural weather backbone into a stable climate emulator.
0.4 Mparameters wrapped around the frozen backbone. NeurIPS 2026 Tackling Climate Change with ML workshop.
Global wildfire forecasts at 0.25° that use pixel-level fire memory and a recency-weighted climatology.
0.724 to 0.642AUPRC from 8-day to 128-day lead on SeasFire. The reproduced TeleViT baseline scores 0.624 to 0.607.
Lead-aware probabilistic streamflow forecasts that bridge the reanalysis-to-forecast gap.
0.655 / 0.550 / 0.481median KGE′ at 1, 5 and 10 days over about 2,100 stations. The AIFL-recipe baseline scores 0.618 / 0.459 / 0.340.
Learned downscaling from 5.625° to 1.5° on 69 variables, with an exact consistency projection onto the coarse input.
1.83 Mparameters. The output reproduces its coarse input to about 1e-7.
Sparse autoencoders built from B-spline units that open up Sonny's internal features.
Feature atlas, case studies and causal steering.
Low-rank Kolmogorov-Arnold generator flows for interpretable operator learning, with closed-form distillation of what was learned.
3.4 to 11×lower out-of-distribution loss than the Kolmogorov-Arnold neural operator on synthetic benchmarks.