Aeromancer
Physics-aware AI that turns climate fields into readable insights
Building a modeling and inference platform that learns spatiotemporal norms in atmospheric processes to detect, attribute, and narrate anomalous climate evolution
ERA5 (6-hourly)VideoMAE backboneResidual modeling (NeuralGCM)Event graphs → summariesAnnotated video outputs
How it works
From fields → events → narrative
- Ingest reanalysis/forecast variables (e.g., MSLP, Z500, vorticity)
- Detect & track evolving systems (genesis → peak → decay)
- Generate event timelines, annotated clips, and short reports
Outputs
Built for scientific narration
- Event JSON (location, intensity, motion, confidence)
- Video composites (overlays + captions)
- Technical + plain-language summaries
Status
Prototype in progress
Current focus: cyclone-centered clips on ERA5, evaluation harness, and residual-informed anomaly scoring. Next: deploy the inference worker and ship a minimal API + docs.