## 1. Field Overview Recent work in **deep learning for extreme-weather-relevant climate and weather prediction** is dominated by large-scale, data-driven weather prediction models trained on reanalysis or operational weather data. The mainstream routes are: **autoregressive global forecasting models** using graph neural networks or Transformer-like architectures; **3D Earth-specific neural networks** for medium-range prediction; **probabilistic diffusion / generative ensemble models** for uncertainty and extremes; **hybrid physics–ML general circulation models** for weather-to-climate simulation; and **regional high-resolution downscaling or stretched-grid models** for precipitation extremes and atmospheric rivers. Benchmarking is increasingly standardized through frameworks such as WeatherBench 2, but there is still no universal ranking across all extreme event types, lead times, regions, and variables. ([PubMed][1]) **Selection basis:** The list below is based on influence, methodological representativeness, reported benchmark strength, relevance to extreme-event prediction, and publication/source credibility, rather than a universal “best model” ranking. --- ## 2. Selected Representative Papers ### 1. **Learning Skillful Medium-Range Global Weather Forecasting / GraphCast** * **Publication Year:** 2023 * **Venue or Source:** *Science* * **Source Link:** ([PubMed][1]) * **Core Contribution:** Introduced **GraphCast**, a graph neural network weather model that predicts hundreds of variables for 10-day global forecasts at 0.25° resolution and reports improved severe-event prediction for tropical cyclones, atmospheric rivers, and extreme temperatures. ([PubMed][1]) ### 2. **Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks / Pangu-Weather** * **Publication Year:** 2023 * **Venue or Source:** *Nature* * **Source Link:** ([Nature][2]) * **Core Contribution:** Proposed **Pangu-Weather**, using a 3D Earth-specific Transformer architecture and hierarchical temporal aggregation to produce fast global medium-range forecasts, with explicit discussion of extreme-event forecasting and model-smoothing limitations. ([Nature][2]) ### 3. **The Operational Medium-Range Deterministic Weather Forecasting Can Be Extended Beyond a 10-Day Lead Time / FengWu** * **Publication Year:** 2025 * **Venue or Source:** *Communications Earth & Environment* * **Source Link:** ([Nature][3]) * **Core Contribution:** Presented **FengWu**, a multimodal, multitask Transformer-based forecasting system with variable-specific encoders/decoders, cross-modal fusion, uncertainty loss, and replay-buffer training to extend useful medium-range forecasting beyond 10 days. ([Nature][3]) ### 4. **Scaling Transformer Neural Networks for Skillful and Reliable Medium-Range Weather Forecasting / Stormer** * **Publication Year:** 2024 * **Venue or Source:** NeurIPS 2024 / arXiv * **Source Link:** ([arXiv][4]) * **Core Contribution:** Introduced **Stormer**, a relatively simple Transformer backbone with weather-specific embedding, randomized dynamics forecasting, and pressure-weighted loss, showing competitive WeatherBench 2 performance with less training data and compute. ([arXiv][4]) ### 5. **Probabilistic Weather Forecasting with Machine Learning / GenCast** * **Publication Year:** 2025 * **Venue or Source:** *Nature* * **Source Link:** ([Nature][5]) * **Core Contribution:** Introduced **GenCast**, a probabilistic ML weather model that generates stochastic 15-day ensemble forecasts and reports better performance than ECMWF ENS on 97.2% of evaluated targets, including extreme weather and tropical cyclone tracks. ([Nature][5]) ### 6. **Neural General Circulation Models for Weather and Climate / NeuralGCM** * **Publication Year:** 2024 * **Venue or Source:** *Nature* * **Source Link:** ([Nature][6]) * **Core Contribution:** Proposed **NeuralGCM**, a hybrid differentiable general circulation model combining a physical dynamical core with learned neural parameterizations, enabling weather forecasts and multi-decade climate simulations with large computational savings. ([Nature][6]) ### 7. **A Foundation Model for the Earth System / Aurora** * **Publication Year:** 2025 * **Venue or Source:** *Nature* * **Source Link:** ([Nature][7]) * **Core Contribution:** Introduced **Aurora**, a 1.3B-parameter Earth-system foundation model using a 3D Swin Transformer processor and Perceiver-style encoder/decoder, fine-tuned for high-resolution weather, tropical cyclone tracks, air pollution, and ocean waves. ([Nature][7]) ### 8. **End-to-End Data-Driven Weather Prediction / Aardvark Weather** * **Publication Year:** 2025 * **Venue or Source:** *Nature* * **Source Link:** ([Nature][8]) * **Core Contribution:** Presented **Aardvark Weather**, an end-to-end system that ingests observations directly and produces global gridded and local station forecasts, reducing dependence on conventional NWP initialization and post-processing. ([Nature][8]) ### 9. **Residual Corrective Diffusion Modeling for km-Scale Atmospheric Downscaling / CorrDiff** * **Publication Year:** 2025 * **Venue or Source:** *Communications Earth & Environment* * **Source Link:** ([Nature][9]) * **Core Contribution:** Proposed a two-stage residual diffusion downscaling approach that maps 25 km weather inputs to 2 km regional fields, sharpening fronts, intensifying typhoons, and synthesizing rainbands for hazard-relevant high-resolution prediction. ([Nature][9]) ### 10. **A Regional High Resolution AI Weather Model for the Prediction of Atmospheric Rivers and Extreme Precipitation** * **Publication Year:** 2025 * **Venue or Source:** *npj Climate and Atmospheric Science* * **Source Link:** ([Nature][10]) * **Core Contribution:** Developed a stretched-grid, graph-transformer AI weather model with 6 km resolution over the Western United States, targeting atmospheric rivers and extreme precipitation events that coarse global models often underestimate. ([Nature][10]) --- ## 3. Comparison Summary **Dominant method family:** Transformer-derived and autoregressive global forecasting models are the dominant architecture family. GraphCast uses graph neural networks, Pangu-Weather uses 3D Earth-specific networks, FengWu and Stormer use Transformer-centered designs, and Aurora extends this trend into foundation-model territory. ([PubMed][1]) **Accuracy and benchmark performance:** GraphCast, Pangu-Weather, FengWu, Stormer, GenCast, and Aurora are most directly positioned around forecast skill and benchmark superiority. WeatherBench 2 is especially important because it provides an open framework for comparing physical and data-driven global weather models with standardized metrics. ([PubMed][1]) **Uncertainty, robustness, and real-world deployment:** GenCast and CorrDiff represent the generative/probabilistic direction, which is important for extreme events because decision-making requires uncertainty estimates rather than only deterministic means. NeuralGCM addresses physical consistency and longer climate simulation, Aurora emphasizes foundation-model transfer across Earth-system tasks, and Aardvark targets end-to-end operational simplification from observations to forecasts. ([Nature][5]) **Key research gaps:** The main gaps are reliable prediction of **rare, high-intensity extremes**, calibrated uncertainty at local scales, precipitation realism, and generalization to climates outside the training distribution. Pangu-Weather explicitly notes that AI forecasts can become overly smooth and risk underestimating extreme-event magnitudes, while NeuralGCM reports limitations when extrapolating to substantially different future climates. ([Nature][2]) **Promising future directions:** The strongest directions are high-resolution regional AI models for precipitation and atmospheric rivers, probabilistic diffusion ensembles, hybrid physics–ML climate models, and end-to-end systems that can be optimized for specific hazards, regions, and decision contexts. [1]: https://pubmed.ncbi.nlm.nih.gov/37962497/ "Learning skillful medium-range global weather forecasting - PubMed" [2]: https://www.nature.com/articles/s41586-023-06185-3 "Accurate medium-range global weather forecasting with 3D neural networks | Nature" [3]: https://www.nature.com/articles/s43247-025-02502-y "The operational medium-range deterministic weather forecasting can be extended beyond a 10-day lead time | Communications Earth & Environment" [4]: https://arxiv.org/abs/2312.03876 "[2312.03876] Scaling transformer neural networks for skillful and reliable medium-range weather forecasting" [5]: https://www.nature.com/articles/s41586-024-08252-9 "Probabilistic weather forecasting with machine learning | Nature" [6]: https://www.nature.com/articles/s41586-024-07744-y "Neural general circulation models for weather and climate | Nature" [7]: https://www.nature.com/articles/s41586-025-09005-y "A foundation model for the Earth system | Nature" [8]: https://www.nature.com/articles/s41586-025-08897-0 "End-to-end data-driven weather prediction | Nature" [9]: https://www.nature.com/articles/s43247-025-02042-5 "Residual corrective diffusion modeling for km-scale atmospheric downscaling | Communications Earth & Environment" [10]: https://www.nature.com/articles/s41612-025-01265-9 "A regional high resolution AI weather model for the prediction of atmospheric rivers and extreme precipitation | npj Climate and Atmospheric Science"