U-Cast (Rose-STL-Lab, ICML 2026)
github.com/rose-stl-lab/u-castSurprisingly simple and efficient frontier probabilistic weather forecaster built on a standard U-Net trained with deterministic MAE pre-training followed by short CRPS fine-tuning via Monte Carlo Dropout, matching or exceeding the probabilistic skill of GenCast and IFS ENS at 1.5° resolution with >10× less training compute than leading CRPS models and >10× lower inference latency than diffusion models; trains in under 12 H200 GPU-days and generates a 15-day ensemble forecast in 3 seconds, with official code and pretrained checkpoints (UCLA, 41+ stars, Apache 2.0)
Sourced from
- Awesome AI for Science — github.com/rose-stl-lab/u-cast
- GitHub — github.com/rose-stl-lab/u-cast
Related resources
Google DeepMind's diffusion-based ensemble weather forecasting model at 0.25° resolution, outperforming ECMWF ENS on 97.2% of targets up to 15 days ahead, with open-source code and weights (Nature 2024)
Python package for segmenting geospatial data with the Segment Anything Model (SAM), enabling zero-shot object segmentation in satellite and aerial imagery for remote sensing and Earth observation (MIT, 4k+ stars)
High-level open-source geospatial AI package for satellite/aerial imagery analysis, model training, inference, interactive visualization, and QGIS integration, bridging PyTorch/Transformers with remote sensing workflows (MIT, 2026)
Huawei's 3D high-resolution global weather forecast model at 0.25° resolution, first AI method to comprehensively outperform traditional NWP across all variables and lead times, integrated into ECMWF operational forecasts (Nature 2023)
World's first fully open, accelerated weather AI software stack with Medium Range forecasting and Nowcasting models using generative AI (January 2026)
Google Research's hybrid ML/physics atmospheric model combining learned dynamics with physical constraints, outperforming traditional models on 2-15 day forecasts and 40-year climate simulation, developed with ECMWF (Nature 2024)