Models that learned the atmosphere

01Timeline 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

02Systems 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.

03Layer 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.