How microsoft/aurora works
Aurora is an Earth-system forecasting foundation model: users provide a gridded Batch, run a pretrained or task-specialised model, and receive forecasts. The core path covers input representation and normalization, neural encoding and prediction, and autoregressive rollout; additional branches support tracking and Azure AI Foundry job submission and serving. The model’s full forward implementation and several Foundry interactions were not sampled, so edges are limited to relationships shown in excerpts, the source index, or README.
Main entry points: aurora.py Aurora model · encoder.py Perceiver encoder · rollout.py Forecast rollout · batch.py Input batch
Components 22
Forecast workflow
- Input batch
aurora/batch.py - Normalization
aurora/normalisation.py - Aurora model
aurora/model/aurora.py - Forecast rollout
aurora/rollout.py - Forecast output
aurora/batch.py
Model internals
- Perceiver encoder
aurora/model/encoder.py - Perceiver decoder
aurora/model/decoder.py - Swin transformer
aurora/model/swin3d.py - Perceiver blocks
aurora/model/perceiver.py - Fourier embeddings
aurora/model/fourier.py - Level conditioning
aurora/model/levelcond.py - Patch embeddings
aurora/model/patchembed.py - Checkpoint adaptation
aurora/model/compat.py - Insolation features
aurora/insolation.py
Forecast analysis
- Cyclone tracking
aurora/tracker.py
Foundry integration
- Job submission
aurora/foundry/client/api.py - Communication channel
aurora/foundry/common/channel.py - Foundry client
aurora/foundry/client/foundry.py - Serving model
aurora/foundry/common/model.py - MLflow wrapper
aurora/foundry/server/mlflow_wrapper.py
Other components
- Forecast user
- Foundry caller
Connections 27
- Forecast user → Input batch: provides inputs
README.md - Forecast user → Aurora model: runs inference
README.md - Aurora model → Forecast output: returns prediction
README.md - Input batch → Normalization: normalizes with
aurora/batch.py - Aurora model → Input batch: consumes
aurora/model/aurora.py - Aurora model → Perceiver encoder: uses encoder
aurora/model/aurora.py - Aurora model → Perceiver decoder: uses decoder
aurora/model/aurora.py - Aurora model → Swin transformer: uses backbone
aurora/model/aurora.py - Aurora model → Checkpoint adaptation: adapts checkpoints
aurora/model/aurora.py - Aurora model → Insolation features: uses features
aurora/model/aurora.py - Perceiver encoder → Perceiver blocks: uses blocks
aurora/model/encoder.py - Perceiver encoder → Fourier embeddings: uses embeddings
aurora/model/encoder.py - Perceiver encoder → Level conditioning: uses conditioning
aurora/model/encoder.py - Perceiver encoder → Patch embeddings: uses embeddings
aurora/model/encoder.py - Perceiver decoder → Perceiver blocks: uses blocks
aurora/model/decoder.py - Perceiver decoder → Fourier embeddings: uses embeddings
aurora/model/decoder.py - Perceiver decoder → Level conditioning: uses conditioning
aurora/model/decoder.py - Forecast user → Forecast rollout: requests steps
aurora/rollout.py - Forecast rollout → Aurora model: calls forward
aurora/rollout.py - Forecast rollout → Input batch: advances input
aurora/rollout.py - Forecast rollout → Forecast output: yields forecasts
aurora/rollout.py - Foundry caller → Job submission: submits job
aurora/foundry/client/api.py - Job submission → Communication channel: uploads input
aurora/foundry/client/api.py - Job submission → Foundry client: checks status
aurora/foundry/client/api.py - MLflow wrapper → Serving model: wraps serving
package_mlflow.py - Serving model → Forecast rollout: runs forecasts
aurora/foundry/common/model.py - Communication channel → Input batch: serializes batches
aurora/foundry/common/channel.py