README.md

---
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.