| license: apache-2.0 | |
| tags: | |
| - space-safety | |
| - conjunction-assessment | |
| - satellite-collision | |
| - physics-informed | |
| # Panacea — Satellite Collision Avoidance Models | |
| ML models for satellite conjunction assessment, trained on ESA Kelvins CDM dataset. | |
| ## Models | |
| | Model | File | AUC-PR | Description | | |
| |-------|------|--------|-------------| | |
| | Orbital Shell Baseline | `baseline.json` | 0.061 | Altitude-binned collision rates | | |
| | XGBoost | `xgboost.pkl` | 0.988 | Gradient-boosted trees on CDM features | | |
| | PI-TFT | `transformer.pt` | 0.511 | Physics-Informed Temporal Fusion Transformer | | |
| ## Usage | |
| ```python | |
| from huggingface_hub import hf_hub_download | |
| # Download PI-TFT model | |
| path = hf_hub_download("DTanzillo/panacea-models", "transformer.pt") | |
| checkpoint = torch.load(path, map_location="cpu", weights_only=False) | |
| ``` | |
| ## Weekly Fine-Tuning | |
| The PI-TFT model is automatically fine-tuned weekly using Starlink maneuver | |
| detections as proxy training labels. See the | |
| [Panacea repo](https://github.com/DominicTanzillo/Panacea) for details. |