--- title: "Models that learned the atmosphere | Nexus Observatory" description: "In three years, machine-learned weather models went from research papers to operations. They learned from decades of reanalysis…" url: https://observatory.global/stories/ai-earth-models/ publisher: Nexus Observatory Inc. updated: 2026-10-31 --- [Stories](https://observatory.global/stories/)Horizon 2030: AI weather and geospatial foundation models # Models that learned the atmosphere In three years, machine-learned weather models went from research papers to operations12. They learned from decades of reanalysis3, which means they inherit the quality of the observations behind it. **Systems**: Machine-learned forecasting, Geospatial foundation models, Reanalysis **Toward 2030**: Operational AI forecasts, Foundation models for imagery **Figures**: 3, timeline to 2030, systems diagram, layer stack Go to the references **01** Timeline to 2030 ## Three years from paper to operations. A neural operator model in 20224, a 3D transformer and a graph network in 202351, an ensemble diffusion model at the end of 20246, and in February 2025 a data-driven system running operationally at a major forecasting centre72. Figure as text: Machine-learned weather and geospatial models A timeline from 2020 to 2030 with FourCastNet, SatMAE, Pangu-Weather, Prithvi, GraphCast, the AIFS paper, GenCast and operational AIFS. 1. **2020** ERA5 reanalysis described. The hourly global reanalysis most learned models train on. 3 2. **2022-02** Neural operator forecast. A global data-driven model at 0.25°. 4 3. **2022-07** Masked pre-training for satellite imagery. Temporal and multispectral imagery as pre-training data. 8 4. **2023-07** 3D transformer forecasts. Medium-range forecasts with three-dimensional networks. 5 5. **2023-10** Geospatial foundation model. A model pre-trained on harmonised satellite imagery and fine-tuned for floods, burn scars and crops. 9 6. **2023-11** Graph network forecasts. Skill comparable to or better than physics-based systems on most verification targets studied. 1 7. **2024-12** Probabilistic learned ensembles. An ensemble of learned forecasts with calibrated uncertainty. 6 8. **2025-02** Operational AI forecasts. A data-driven global forecast in operations. 2 **02** Systems diagram ## Still built on observations. A learned model forecasts from an initial state, and that state still comes from observations assimilated into an analysis7. Remove the observations and the model has nothing to start from; that dependency runs through the whole chain. Figure as text: Observations, analysis, learned model and users A four-tier diagram: observing systems, data assimilation, the learned and physics models, and forecast users. - **Observations** : Satellites, Stations, Soundings - **Analysis** : Assimilation - **Models** : Learned model, Physics model - **Users** : Forecast services, Local downscaling 8 links. Scenarios: Full observing system; No analysis. **03** Layer stack ## Foundation models read the layers. Geospatial foundation models are pre-trained on large volumes of imagery and then adapted to tasks such as flood mapping, burn scars or crop classification with few labels98. In a stack of layers, they sit between the imagery and the model outputs. Figure as text: Imagery, learned features and task outputs An exploded stack: terrain, imagery, an analysis grid, a learned feature field and a task output with a red threshold line. 1. **Terrain** 2. **Imagery** : Multispectral scenes over time. 3. **Tiles** : The tiles the model reads. 4. **Learned features** : Embeddings from a pre-trained encoder. 5. **Task output** : For example a flood extent; the red line is the decision boundary. 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 Today Weather Imagery Feb 2025 · Weather ### Operational AI forecasts A data-driven global forecast in operations.2 Figure 1 Timeline to 2030 Figure 1. Machine-learned weather and geospatial models. A timeline from 2020 to 2030 with FourCastNet, SatMAE, Pangu-Weather, Prithvi, GraphCast, the AIFS paper, GenCast and operational AIFS. ## References 1. [Learning skillful medium-range global weather forecasting](https://doi.org/10.1126/science.adi2336) . Lam, R. et al., Science 382(6677):1416–1421, 2023. 2. ECMWF’s AI forecasts become operational. European Centre for Medium-Range Weather Forecasts, 2025. 3. [The ERA5 global reanalysis](https://doi.org/10.1002/qj.3803) . Hersbach, H. et al., Quarterly Journal of the Royal Meteorological Society 146:1999–2049, 2020. 4. [FourCastNet: a global data-driven high-resolution weather model using adaptive Fourier neural operators](https://arxiv.org/abs/2202.11214) . Pathak, J. et al., arXiv:2202.11214, 2022. 5. [Accurate medium-range global weather forecasting with 3D neural networks](https://doi.org/10.1038/s41586-023-06185-3) . Bi, K. et al., Nature 619:533–538, 2023. 6. [Probabilistic weather forecasting with machine learning](https://doi.org/10.1038/s41586-024-08252-9) . Price, I. et al., Nature 637:84–90, 2025. 7. [AIFS: ECMWF’s data-driven forecasting system](https://arxiv.org/abs/2406.01465) . Lang, S. et al., arXiv:2406.01465, 2024. 8. [SatMAE: pre-training transformers for temporal and multi-spectral satellite imagery](https://arxiv.org/abs/2207.08051) . Cong, Y. et al., NeurIPS 2022, arXiv:2207.08051, 2022. 9. [Foundation models for generalist geospatial artificial intelligence](https://arxiv.org/abs/2310.18660) . Jakubik, J. et al., arXiv:2310.18660, 2023. ## Related stories [System: End to end **An observation grid, end to end** An observation is not a reading. It is a reading with a place, a time, a calibration and a path that anyone can follow back. An observing grid is the…](https://observatory.global/stories/observation-grid/) [Place: Farmland **Field to forecast** A farm observes water: in the soil, in the crop and in the air above it. Crop water use is computed from weather with a method agronomists have shared for…](https://observatory.global/stories/farmland/) [Horizon 2030: Overview **What changes by 2030** Six shifts will reshape how the Earth is observed before the decade ends: networks that sense, models that forecast, foundation models that read imagery,…](https://observatory.global/stories/horizon-2030/) ## Where this applies [Earth observation & GeoAI](https://observatory.global/earth-observation/)[Data & licensing](https://observatory.global/data-products/)[Software](https://observatory.global/software/)[Edge compute & AI](https://observatory.global/edge-ai/)[Climate](https://observatory.global/industries/climate/)[Agriculture](https://observatory.global/industries/agriculture/)[Insurance and finance](https://observatory.global/industries/insurance/)[Energy](https://observatory.global/industries/energy/)[Field tour: orbit view](https://observatory.global/environments/orbit-view/) © 2026 Nexus Observatory Inc. **Legal** [Privacy notice](https://observatory.global/legal/privacy/) · [Terms of use](https://observatory.global/legal/terms/) · [Cookie policy](https://observatory.global/legal/cookies/) · [Accessibility](https://observatory.global/legal/accessibility/) · [Official languages](https://observatory.global/official-languages/) · [Notices](https://observatory.global/notices/) **Company** [About](https://observatory.global/about/) · [Open governance](https://observatory.global/consortium/) · [Contact](https://observatory.global/contact/)